betaFeedback

WorksPhysics and Astronomy

Neural Network Path Optimization for Finding Transition States on a Machine Learned Potential

Abstract

Transition states govern chemical reaction rates and outcomes, yet locating them remains a central challenge in computational chemistry. Double-ended search methods find transition states given reactant and product structures, but generally rely on discretized path representations developed for density functional theory, where potential energy surface evaluations are expensive. Machine learned interatomic potentials (MLIPs) dramatically reduce this cost and leverage automatic differentiation to access higher-order derivatives efficiently. We introduce Popcornn, a framework to find transition states by optimizing continuous neural network paths atop an MLIP with a parallelized adaptive numerical integrator. This combination eliminates path discretization artifacts, reduces hyperparameter sensitivity, and enables non-local optimization to overcome local features using second-order potential-energy-surface information. Popcornn consistently recovers more correct transition states than standard methods such as nudged elastic band and growing string, across MLIPs on thousands of organic reactions, establishing it as a robust, automation-ready method for high-throughput transition state search. Popcornn's infrastructure is flexible in path objective, integration strategy, and optimization algorithm, making it a practical tool and platform for investigating open questions in MLIP-based path optimization.

*Corresponding author(s). E-mail(s): [email protected]; [email protected]; [email protected];

†These authors contributed equally to this work.

Introduction

The quantum mechanical potential energy surface (PES) describes how an atomistic system’s energy changes during conformational rearrangement and chemical transformations. The most chemically pertinent areas of the PES are stationary points: minima define stable molecules, and first-order saddle points define transition states (TSs) that describe transformations of reactant molecules into product species. Accurately and efficiently locating the TSs on the PES is critical for understanding complex reactive combustion [1, 2], electrolyte decomposition [3, 4, 5, 6, 7], heterogeneous [8] and homogeneous catalysis [9], polymer transformations [10], and more.

PES minima can be reliably found with a variety of different local optimization algorithms using only iterative energy and gradient (1st order and quasi-second order) methods, starting from nearly any reasonable initial guess structure [11, 12]. In contrast, single-ended TS optimization depends heavily on the quality of the initial guess (i.e., how close the structure is to the desired TS), the optimization algorithm employed, and often the need to calculate Hessian (2nd order) information. TS optimization for chemical transformations is typically carried out with density functional theory (DFT). However, when Hessian (2nd order) information is used, DFT eventually becomes intractable for large systems due to quartic scaling with the number of basis functions. As automatically-differentiable machine learned interatomic potentials (MLIPs) progress towards being reliable DFT surrogates, improved accessibility of second-order information has enabled more robust single-ended TS optimization [13, 14, 15, 16].

Even with readily-available Hessians, single-ended TS optimization remains fundamentally constrained by the need for a reasonably good initial guess TS structure [13, 16]. Experts can build such structures by hand, structural templates can be employed for specific reaction classes [17], and generative models can now propose TS structures from reactants and products [18, 19, 20, 21] or from reactants alone [22]. However, these generative models are limited in chemical scope by their TS training data; currently, standard datasets consist of main-group reactions [22, 23]. Therefore, automated initial guess construction across the chemical diversity of complex reaction networks [24, 25] remains an open challenge.

An alternative and more automation-friendly approach is double-ended TS search [12], which locates the TS given reactant and product structures by reformulating the task as a path optimization. It is the natural tool for reaction networks built by enumerating candidate reactant–product pairs [5, 24, 26]. The path can be thought of as an optimizable function xθ(t)x_{\theta}(t) which maps a point on a reaction coordinate or “reaction time” tt to a 3D structure. The path is constrained to return the reactant structure when t=0t=0 and the product structure when t=1t=1: xθ(0)=xRx_{\theta}(0)=x_{\mathrm{R}} and xθ(1)=xPx_{\theta}(1)=x_{\mathrm{P}}. Thus, the transformation of the reactant to the product must necessarily be accomplished along the path. The optimization objective for most double-ended path optimization schemes is derived from or can be interpreted as a path integral:

L=∫01ℓ(x(t)) dt,(1)\mathcal{L}=\int_{0}^{1}\ell(\mathbf{x}(t))\,dt, \tag*{(1)}

where ℓ(x(t))\ell(\mathbf{x}(t)) quantifies the local “path quality” at reaction time tt. In the simplest case, ℓ(x(t))\ell(\mathbf{x}(t)) can be the potential energy E(x(t))E(\mathbf{x}(t)), and the minimization objective LE\mathcal{L}_{E} is the total energy along the path. Decades of work, including the nudged elastic band (NEB) [27, 28] and growing string method (GSM) [29, 30, 31, 32], have been dedicated to obtaining the minimum energy path (MEP) by minimizing LE\mathcal{L}_{E}. The highest energy point along the MEP is guaranteed to be a TS, and the MEP is the gradient descent path from the TS to the reactant and product minima.

Notably, NEB and GSM were developed under the constraints of performing path optimization atop a DFT PES. Despite avoiding prohibitively expensive Hessians, calculating a structure’s total energy and atomic forces with DFT still scales as O(Ne3)O(N_{e}^{3}), becoming computationally challenging or intractable for larger systems. Traditionally, to make the path optimization computationally practical, the path (and its integral Eq. (1)) is discretized to a set of discrete structures (images) such that only tens of DFT calculations are needed at each optimization iteration. However, path discretization has a variety of negative consequences: the potential to miss PES features between discrete images; complexities around how to place images along the path; maintaining path connectivity with fictitious spring forces; the need to treat a “climbing image” differently from the rest of the path; and the introduction of many hyperparameters (related to these details) that often need to be carefully selected by the user for each reaction. All of these consequences are anathema to automation. Perhaps most problematic, path optimization proceeds locally such that each image’s position is updated pursuant only to the energy and forces at that image and its two neighboring images, preventing the path from cohesively mutating into an entirely different track along the PES.

Independent of the path representation, the MEP optimization objective is critically hindered by the MEP’s sensitivity to local structures that do not affect the TS. Although LE\mathcal{L}_{E} only requires PES first derivatives at each optimization iteration and the MEP has attractive mathematical properties [33, 34], it is also sensitive to every part of the path. This sensitivity is a crucial drawback, making the optimization landscape substantially more complex and difficult to navigate. In practice, the primary goal of path optimization should be to locate the TS. With the TS in hand, we can then obtain the MEP and/or confirm the associated endpoints via intrinsic reaction coordinate (IRC) or quasi-IRC [33, 35] calculations. However, directly targeting the TS – rather than the MEP – requires a force-based path objective [36]; but a force-based path objective necessitates the calculation of second derivatives along the path at each iteration. While intractable with DFT, this presents an opportunity for improved path optimization methodology atop differentiable MLIPs.

In this work, we present a software platform for MLIP-based double-ended TS search that we call Popcornn: “Path Optimization with a Continuous Representation Neural Network.” Popcornn leverages a neural network to parameterize the continuous path from reactant to product, combined with a GPU-parallelized adaptive numerical path integrator to calculate the path objective – a critical combination we find necessary to ensure comprehensive PES sampling and to avoid many issues of traditional path discretization. This combination further facilitates global and cohesive path optimization by differentiating the path objective through the MLIP with respect to the path’s weights and biases. By default, Popcornn employs a force-based optimization objective that directly targets the TS rather than seeking the MEP. We demonstrate that, with fixed default settings, Popcornn consistently achieves improved robustness and recovery of intended TSs, when compared with NEB and GSM, across multiple MLIPs and thousands of in-distribution and out-of-distribution test reactions. Additionally, Popcornn is engineered to be flexible with respect to path objective, numerical integration strategy, optimization algorithm, and path initialization strategy. Thus, it is both of immediate practical utility for double-ended TS search atop a pre-trained MLIP, and it is also envisioned as a powerful platform for subsequent investigation and further improvement of MLIP-based path optimization methodologies.

Results

Popcornn framework

Popcornn uses a simple fully-connected neural network to parameterize the path,

xθ(t)=(1−t)xR+txP⏟linear prior+t(1−t)fθ(t)⏟perturbation(2)x_{\theta}(t) = \underbrace{(1-t)x_{\mathrm{R}} + tx_{\mathrm{P}}}_{\text{linear prior}} + \underbrace{t(1-t)f_{\theta}(t)}_{\text{perturbation}} \tag*{(2)}

where t∈[0,1]t \in[0,1] is the reaction time, xRx_{\mathrm{R}} and xPx_{\mathrm{P}} are the reactant and product endpoints, and fθ(t)f_{\theta}(t) is a fully-connected neural network with adjustable parameters θ\theta (Figure 1a). The path (xθ(t)∈R3N)(x_{\theta}(t) \in\mathbb{R}^{3N}) and neural network (fθ(t)∈R3N)(f_{\theta}(t) \in\mathbb{R}^{3N}) model the positions of all NN atoms. The path formulation in Eq. (2) employs a perturbation from a linear interpolation between xRx_{\mathrm{R}} and xPx_{\mathrm{P}}, while enforcing hard endpoint constraints: xθ(0)=xRx_{\theta}(0) = x_{\mathrm{R}} and xθ(1)=xPx_{\theta}(1) = x_{\mathrm{P}}. We observe that the perturbative correction to a linear prior is important for path quality and optimization speed; see Supplementary Note A for details. This path formulation was developed independently and concurrently in references [37, 38, 39]; additionally, neural path models have also been trained across many reactions to predict paths directly, from the endpoints [39] or outward from a TS [22], whereas Popcornn optimizes each path on the target PES. The continuous path allows one to probe the molecular structure at any point, guaranteeing no discontinuities and allowing all relevant PES features along the path to be sampled during optimization. Critically, we note that changing any internal neuron parameter in xθ(t)x_\theta(t) modifies all atomic positions across the entire reaction path, enabling non-local path mutation rather than local bead perturbations.

Schematic illustration of Popcornn

Figure 1. Schematic illustration of Popcornn. (a) The continuous reaction path is parameterized by a simple neural network. (b) The path objective of choice is calculated via a GPU-parallelized adaptive numerical path integrator. (c) The gradient of the objective with respect to the path neural network parameters is obtained via backpropagation through the differentiable potential, enabling path optimization via gradient descent.

To realize our continuous formulation’s (Eqs. (1) and (2)) full capabilities, the loss must also be evaluated via adaptive numerical integration, rather than a coarse or other rigid summation scheme. Thus, unlike conventional methods that approximate the loss (Eq. (1)) using discrete points, we evaluate the path integral with the GPU-parallelized adaptive numerical path integrator PadaQuad [40] (Figure 1b). PadaQuad globally updates the entire path by integrating the gradient of the loss with respect to all network parameters along the entire path (Eq. (5)), driving cohesive path optimization via gradient descent (Figure 1c). The adaptivity of PadaQuad is central to avoiding fixed discretization, and it focuses PES evaluations in regions of high variation and large contributions to the integral while limiting evaluations in slowly varying regions (Supplementary Note E). The adaptive numerical integration of Eq. (1) precludes the need for fictitious spring force terms. Instead, PadaQuad does this by adaptively adding, removing, and adjusting evaluation points based on the integral error, independent of the PES value. This adaptivity places evaluations where needed, removing the need for spring forces to affix them and creating an efficient and error-bound evaluation of the path. We provide more details on PadaQuad in Supplementary Note E.

We explore two different path objectives: one which targets the minimum energy path, and the other which directly targets the TS via the path which goes least high on the PES. Crehuet and Bofill [41] introduced the variational reaction energy (VRE) path objective

LVRE=∫01∥F(xθ(t))∥∥vθ(t)∥ dt,(3)\mathcal{L}_{\mathrm{VRE}} = \int_{0}^{1} \lVert\mathbf{F}(\mathbf{x}_{\theta}(t)) \rVert\lVert\mathbf{v}_{\theta}(t) \rVert\, \mathrm{d}t, \tag*{(3)}

where F(xθ(t))∈R3N\mathbf{F}(\mathbf{x}_{\theta}(t)) \in\mathbb{R}^{3N} is the atomic forces at positions xθ(t)\mathbf{x}_{\theta}(t), and vθ(t)=dxθ(t)/dt∈R3N\mathbf{v}_{\theta}(t) = \mathrm{d}\mathbf{x}_{\theta}(t)/\mathrm{d}t \in\mathbb{R}^{3N} is the tangent to the path. Minimizing LVRE\mathcal{L}_{\mathrm{VRE}} results in a minimum energy path, also known as a gradient descent path [42, 43]. This MEP is the goal of NEB and GSM. Birkholz and Schlegel [36] introduced the projected VRE (pVRE) loss as a variational bound of LVRE\mathcal{L}_{\mathrm{VRE}}

LpVRE=∫01∣F(xθ(t))⋅vθ(t)∣ dt=∑α2Eα,TS−Eα,R−Eα,P,(4)\mathcal{L}_{\mathrm{pVRE}} = \int_{0}^{1} \left| \mathbf{F}(\mathbf{x}_{\theta}(t)) \cdot\mathbf{v}_{\theta}(t) \right| \, \mathrm{d}t = \sum_{\alpha} 2E_{\alpha,\mathrm{TS}} - E_{\alpha,\mathrm{R}} - E_{\alpha,\mathrm{P}}, \tag*{(4)}

which is sensitive to the angle between F(xθ(t))\mathbf{F}(\mathbf{x}_{\theta}(t)) and vθ(t)\mathbf{v}_{\theta}(t), and which can be thought of as a sum over all activation barriers α\alpha. Minimizing LpVRE\mathcal{L}_{\mathrm{pVRE}} results in a path that has the lowest PES barrier, thereby directly targeting the TS while being insensitive to structures along the rest of the path. Later, we show LpVRE\mathcal{L}_{\mathrm{pVRE}} produces qualitatively simpler and more robust paths.

The path xθ(t)\mathbf{x}_{\theta}(t) is updated by the gradient of the path integral loss given by

∂L∂θ=∫01∂ℓ(t)∂θ dt=∫01∂ℓ(xθ(t))∂xθ∂xθ(t)∂θ dt.(5)\frac{\partial\mathcal{L}}{\partial\theta} = \int_{0}^{1} \frac{\partial\ell(t)}{\partial\theta} \, \mathrm{d}t = \int_{0}^{1} \frac{\partial\ell(\mathbf{x}_{\theta}(t))}{\partial\mathbf{x}_{\theta}} \frac{\partial\mathbf{x}_{\theta}(t)}{\partial\theta} \, \mathrm{d}t. \tag*{(5)}

Numerically, this gradient can be calculated in two ways: taking the gradient of the completed integral; or integrating the gradient of the integrand. In the former case, the evaluation points are determined by the loss integral since the gradient is taken after. In the latter case, the gradient is more accurate because the evaluation points are chosen to minimize the error of the gradient by placing them where the gradient of the loss varies the most. We use the latter formulation, though Popcornn and PadaQuad support both options. The gradient is one order higher than the highest order derivative in the quantity ℓ\ell. That is, if the integrand ℓ\ell contains only the potential energy (like in NEB), only the first derivative of the PES (atomic forces) would be needed. On the other hand, if ℓ\ell involves the forces (VRE and pVRE), then the second derivative (the Hessian) is necessary for the path differentiation. The steep computational cost of DFT-based second derivatives has prevented force-based path objectives from being realistically viable, until the recent advent of differentiable MLIPs that are sufficiently accurate to describe reactive PESs for chemical systems of interest [13].

The LpVRE\mathcal{L}_{\mathrm{pVRE}} loss admits a natural force-based convergence criterion based on the optimization gradient Eq. (5). Established energy-minimization methods like NEB and GSM leverage modified forces (perpendicular gradients on the images and spring forces) to determine the convergence of the optimization [28]. The LpVRE\mathcal{L}_{\mathrm{pVRE}} loss isolates and targets the TS (Supplementary Note B.2), while removing intermediate extrema along xθ(t)\mathbf{x}_{\theta}(t). Specifically, the gradient of LpVRE\mathcal{L}_{\mathrm{pVRE}} becomes proportional to the TS force as the path converges (Eq. (47)). Therefore, the gradient of LpVRE\mathcal{L}_{\mathrm{pVRE}} is a physically consistent counterpart to convergence criteria used by the aforementioned established methods. Given this correspondence, we set the same convergence criteria for Popcornn, NEB, and GSM for the following chemical reaction evaluations.

To understand the differences between paths optimized with the LpVRE\mathcal{L}_{\mathrm{pVRE}} and LVRE\mathcal{L}_{\mathrm{VRE}} losses, we demonstrate Popcornn with both loss functions on the model Müller-Brown potential, revealing the advantages of LpVRE\mathcal{L}_{\mathrm{pVRE}}. Figure 2 shows the Popcornn optimization of xθ(t)x_{\theta}(t) on the Müller-Brown potential for the LVRE\mathcal{L}_{\mathrm{VRE}} and LpVRE\mathcal{L}_{\mathrm{pVRE}} losses. We observe xθ(t)x_{\theta}(t) optimized with LpVRE\mathcal{L}_{\mathrm{pVRE}} converges to the TS roughly three times faster than with LVRE\mathcal{L}_{\mathrm{VRE}}. The LpVRE\mathcal{L}_{\mathrm{pVRE}} optimization quickly locates the TS and converges to a stable path, while the LVRE\mathcal{L}_{\mathrm{VRE}} optimization struggles to navigate the PES within the first 200 steps due to LVRE\mathcal{L}_{\mathrm{VRE}}’s more complex locally-aware loss landscape. Given that our stated goal is to locate the TS as robustly as possible, the TS-focused LpVRE\mathcal{L}_{\mathrm{pVRE}} is the superior path objective. For the remainder of this work, we use LpVRE\mathcal{L}_{\mathrm{pVRE}} as our optimization objective unless otherwise stated. We provide more detailed discussion and intuition of the LVRE\mathcal{L}_{\mathrm{VRE}} and LpVRE\mathcal{L}_{\mathrm{pVRE}} losses in Supplementary Note B.

Figure 2. Different loss functions on the Muller-Brown potential. (a) The VRE loss tries to find the minimum energy path, which can hinder efficient TS finding. (b) The pVRE loss focuses on minimizing the barrier height, directly targeting the TS. (c) The pVRE loss fully converges within 150 steps, whereas the VRE loss takes 500 steps to converge.

It is well established that the quality of the initial path provided for path optimization dramatically impacts the success rate of the path optimization task and the likelihood of recovering the intended TS [44]. In the molecular context, the geodesic method of Zhu et al. [45] and the IDPP method [46] are commonly used. Both leverage discretized path optimizations with custom objectives atop bespoke analytic pairwise potentials to improve upon a naive linear path. This avoids steric clashes that could yield critically problematic DFT convergence or erroneous MLIP predictions. Popcornn extends its continuous path representation, adaptive numerical integration, and backpropagation-optimization machinery to initial path construction. Similar to MLIPs, one can add new path initializations to the Popcornn ecosystem [47]. We implemented a continuous path geodesic optimization atop the pairwise potential given by Zhu et al. [45], and we refer to this approach as Popcornn-geodesic. As we show below, this substantially improves path optimization outcomes.

Chemical reaction evaluations

To evaluate Popcornn’s ability to robustly locate intended transition states in a systematic manner, we conduct large-scale experiments across different MLIPs on both in-distribution and out-of-distribution reactions. By leveraging reaction test sets containing thousands of examples, we seek to ensure that we are drawing reliable, statistically significant conclusions, while demonstrating Popcornn’s robustness and high-throughput capabilities. We find the reaction path in three steps: we first optimize the endpoint structures on the MLIP (Figure 3a); secondly, we construct an initial path (between the optimized endpoints) atop a pairwise potential using discretized geodesic or Popcornn-geodesic (Figure 3b); finally, we optimize the path atop the MLIP (Figure 3c). To find the TS, we apply a TS optimization method to the structure along the optimized path with maximal energy and minimal forces. To validate the proposed TSs, we apply (quasi)IRC calculations to recover endpoints (Figure 3d) that are then compared with initial endpoints (Figure 3e). Experimental success, as measured by graph matching between initial endpoints and IRC endpoints, is categorized by a quality hierarchy:

Figure 3. Systematic evaluation of double-ended TS finding methods. (a) The endpoints are individually optimized to PES minima. (b) An initial path connecting the initial optimized endpoints is constructed via optimization atop a pairwise potential to approximate the actual reaction path and resolve steric clashes (Geodesic, Popcornn-geodesic). (c) Path optimization is performed on the MLIP PES, starting from the initial path from (b) for NEB and Popcornn, or directly from the optimized endpoints for GSM. (d) The highest point on the optimized path undergoes a single-ended TS optimization and is connected to the reaction endpoints using an IRC. (e) The IRC endpoints are compared with the initial endpoints to check how well the optimized TS corresponds to the intended reaction TS.

  • Intended: Both IRC endpoints match both initial endpoints.

  • Partially intended: One IRC endpoint matches one initial endpoint.

  • Unintended: Neither IRC endpoint matches either initial endpoint; however, the IRC endpoints also do not match each other, i.e., a reaction has still occurred.

  • No reaction: IRC endpoints match, i.e., the transition state found corresponds only to a conformational change, not a reaction.

We benchmark Popcornn against the two most common path optimization methods, NEB [27, 28] and GSM [29, 30, 31, 32]. For NEB, we compare two different initial path construction strategies: discretized geodesic and Popcornn-geodesic. For GSM, initial path construction is not required; the path is built concurrently with optimization. For Popcornn, Popcornn-geodesic constructs the initial path prior to Popcornn optimization atop the MLIP with the pVRE optimization objective. We note that for each method we run all experiments with multiple different sets of hyperparameters and present results from only the single best-performing set. This procedure is most representative of performance in a simple high-throughput context, yielding a fair comparison between methods. Further discussion of NEB hyperparameters is provided in § 4.4. Supplementary Note F presents ablation tests across bead density and path initialization methods for NEB.

We perform experiments on small organic reactions with MLIPs trained on Transition-1x [23] (T1x), and we test on out-of-distribution (OOD) reactions present in the Grambow-Green dataset (GG) [48] but absent in T1x. The reactions in T1x are derived from GG, but in some cases DFT NEB fails to find the path of the GG reaction, and that reaction is excluded from T1x. These challenging 1,342 reactions where DFT NEB fails make up our OOD test dataset. When comparing against NEB, we use standard settings of 10 and 20 beads, along with an exhaustive 100 bead optimization. The 100 bead optimization roughly matches the number of Popcornn evaluations, thus extending NEB towards the numerical integration regime for a more direct comparison with Popcornn. We additionally tested NEB with 1000 beads, but we found significantly worse behavior.

Our first experiment compares Popcornn and NEB performance on small organic reactions using the NewtonNet MLIP [49] trained on T1x [23] from Ref. [13]. We validate each test reaction as having a true TS on the NewtonNet PES by running single-ended TS optimizations from the corresponding TS structure in the Grambow-Green dataset, followed by IRC calculations that must recover endpoints with the intended bonding. This ensures that any result less than “Intended” is solely due to a failure of the double-ended TS search procedure, and not incompleteness of the MLIP energy surface.

Popcornn-geodesic initial path construction followed by Popcornn-pVRE atop the MLIP yields superior TS recovery over NEB with either initial path construction procedure. Popcornn achieves 81% “Intended” TSs (Figure 4b, right bar) compared with 44%, 65%, and 53% for discrete geodesic →\to NEB with 10, 20, and 100 beads, respectively (Figure 4b, left bars). Interestingly, the Popcornn-geodesic initial path improves over the discrete geodesic by 15% for 100 NEB beads, and slightly worsens for 10 and 20 beads. This indicates that a continuous or many-bead path benefits significantly from continuous optimization of the initial path. In contrast, starkly discretized path

Popcornn vs. NEB on small organic reactions using NewtonNet trained on Transition-1x

Figure 4. Popcornn vs. NEB on small organic reactions using NewtonNet trained on Transition-1x. (a) An example of the out-of-distribution reactions, those in the Grambow-Green dataset, on which we are testing. (b) Popcornn finds more “Intended” TSs across various NEB setups with different initializations (Geodesic and Popcornn-Geodesic) and number of beads (10, 20, 100). The number at the bottom of each bar is the average time, in seconds, for the TS optimization to converge. (c) Popcornn’s convergence rate to correct TSs is competitive with low-bead NEB for short runtimes, and ultimately outperforms all NEB variants.

optimizations are not significantly impacted by a continuous initial path. In addition to being more robust at convergence, Popcornn is comparable in speed to both 10 and 20 bead NEB optimizations within the same accuracy range: roughly aligning with the lighter dashed blue lines before 200 s (Figure 4c). Popcornn is significantly faster than the 100 bead path optimization. Further comparison of alternative initial path construction strategies for NEB is provided in Supplementary Note F.

The second experiment compares Popcornn and GSM performance on organic reactions using two popular MLIPs: LeftNet [51] and MACE [52, 53]. Each MLIP was trained or fine-tuned on T1x by Zhao et al. [50], and we follow the training/validation split of T1x used by Zhao et al. for the set of test reactions. Rather than confirming the presence of a TS corresponding to each reaction on each separate MLIP PES, they confirmed the presence of the corresponding TS on the DFT PES used for MLIP training, yielding 960 DFT-validated test reactions. Therefore, it is possible that a failure to identify an “Intended” here may be due to a failure of the double-ended TS search procedure or because the TS does not exist on the MLIP PES. Zhao et al. [50] showed that energy and force mean absolute errors were best for LeftNet and slightly worse for MACE.

Popcornn yields superior TS recovery over GSM across both MLIPs (Figure 5b). Popcornn achieves 84% and 85% “Intended” TSs compared with 74% and 69% for GSM with LeftNet and MACE, respectively. While the absolute success of GSM falls from LeftNet to MACE, corresponding to decreasing accuracy of the MLIP, Popcornn’s success rate remains essentially unchanged. This demonstrates how Popcornn’s numerical integration increases the robustness of the optimized path and TS by disproportionately improving upon less accurate MLIPs. We further observe that the GSM “Intended” rate is weakly affected by the number of beads, although the convergence time increases faster with the number of beads for MACE than it does for LeftNet. Popcornn is faster than 25-bead GSM, but roughly 2 to 3 times slower than 9-bead GSM.

Figure 5. Popcornn vs. GSM on organic reactions using LeftNet and MACE trained or fine-tuned on Transition-1x by Zhao et al [50]. (a) An example of the test reactions evaluated here that are in T1x but not used in training. (b) Popcornn improves recovery of “Intended” TSs for test reactions by 10% and 16% for LeftNet and MACE, respectively, and across varying numbers of beads (9 and 25). The number at the bottom of each bar is the average time, in seconds, for the TS optimization to converge.

Discussion

Popcornn is a robust double-ended TS search framework explicitly designed to leverage the distinct advantages of MLIPs. MLIP efficiency and differentiability via simple backpropagation allow us to address fundamental weaknesses in previous approaches by employing a continuous path representation coupled with adaptive numerical integration of a path objective that directly targets the TS rather than the entire MEP. Further, the batched parallelization of our adaptive integrator allows Popcornn to take full advantage of modern GPU hardware. In contrast to typical small-scale comparisons of path optimization methods, we demonstrate Popcornn’s advantages via high-throughput evaluations with thousands of reactions at various levels of difficulty. While we evaluate robustness and computational runtime, we consider robustness to be the dominant factor that should govern method selection in the MLIP context. This is because probing the PES will become extremely fast/cheap with future ML hardware and software advancements. Popcornn consistently outperforms NEB and GSM across multiple MLIPs and reaction sets while simultaneously reducing the number of hyperparameters that a user must choose, making it well-suited for high-throughput double-ended TS search.

We attribute much of Popcornn’s success to the symbiotic combination of a continuous path representation xθ(t)x_{\theta}(t) and an adaptive numerical integrator. The importance of incorporating careful numerical considerations within scientific machine learning methods has been highlighted previously [54, 55, 56, 57, 58, 59]. Although properly integrating Eq. (1) requires a continuous path, the continuous path alone is not sufficient. Earlier work by Ramakrishnan et al. [38] employ neural network paths with fixed and discrete integral approximation schemes that are still vulnerable to missing essential PES structure along the path. Popcornn, instead, leverages the continuous nature of the path to enable adaptive numerical integration. We therefore stress the importance of simultaneously having a continuous path xθ(t)x_{\theta}(t) and an accurate numerical integration scheme built into the path optimization.

Popcornn is a highly flexible framework built to adapt to and grow with advancing machine learning, chemistry, and optimization techniques. Specifically, by modularizing the path representation, PES, path integrator, and path objective, a user can swap any functionality for their own method. Various MLIPs, including Newton-Net [13, 49], MACE [52, 53], ANI-1x [60, 61], LeftNet [51], Orb [62], and UMA [63], are already supported, and new models can be easily implemented. While we used vanilla neural networks for the path representation in this work, spline representations are also available, and more sophisticated architectures are possible. Similarly, the objective integrator and the objective function can be replaced. The modular architecture empowers chemists to systematically investigate how path representation, objective, optimization, integration, and PES calculations interact; an open methodological question that has thus far been difficult to address with monolithic implementations. We envision the Popcornn framework as a powerful software framework for subsequent path optimization methodology development as well as for establishing additional diverse reaction evaluations.

While the Popcornn framework and the test systems used here are all molecular and non-periodic, we are currently generalizing Popcornn to periodic systems, allowing it to be employed for double-ended transition state optimization for catalysis and material systems. For such periodic non-molecular systems, future work will determine if the geodesic loss and the pairwise potential from Zhu et al. [45] (designed for small molecules) will be sufficient, or if a new interpolation loss (e.g., a continuous implementation of IDPP) will be needed.

Reaction networks are broadly constructed in one of two ways [25]. Exploration-based construction grows the network outward from the reactants with a single-ended TS search, discovering rather than presupposing products [22, 64]. To avoid combinatorial blowup in the number of reactions, one must constrain which transformations are sampled and which species are expanded as the network deepens. Enumeration-based construction instead bounds the chemical space up front by enumerating candidate species and reactant–product pairs [5, 24, 26]. Each pair then requires a double-ended TS search, which imposes no restriction on the bond-change pattern connecting the endpoints. Popcornn is designed for this second setting. It is also complementary to generative TS models [18, 19, 20, 21, 22], which can rapidly propose accurate TSs within the chemistry represented in their training data but require curated sets of converged TSs to train. Since Popcornn is a tool that consumes MLIPs, its scope is only set by the underlying pre-trained MLIP, rather than any one particular dataset coverage. For example, Popcornn already works with foundation MLIPs trained on OMol25 [63, 65] already span transition metals and diverse charge and spin states, chemistry for which reactive TS data remain scarce.

Finally, as MLIPs continue to improve in speed, accuracy, and generality [66], we believe that robust transition state search via Popcornn and subsequent improvements atop our software framework will enable the construction of massive chemical reaction networks, allowing us to understand and eventually design complex reactive processes in electrolytes, catalysis, polymers, and beyond.

Methods

Interfacing with Popcornn

Here, we describe Popcornn in greater detail, walk through important design choices, and outline key steps in applying Popcornn. This section generally follows the Popcornn workflow and highlights key components that a user may wish to adapt to their specific problem. Supplementary Note G gives pseudocode for each step, and Algorithm 1 summarizes a complete run.

Input Endpoints into Popcornn

The Popcornn package takes optimized endpoints as inputs. Popcornn itself does not optimize the endpoints and expects the user to have done so with their preferred method. In this work, the endpoints are first relaxed with BFGS on the MLIP PES. Popcornn fixes the input endpoints in Cartesian coordinates, thus requiring the endpoints to be translationally and rotationally aligned. See Supplementary Note A for more discussion.

The Path Model

A single 2-layer perceptron is used for the path representation, following Eq. (2). Concretely, the neural network fθf_{\theta} is Linear⁡(1,128)→GELU⁡()→Linear⁡(128,3N)\operatorname{Linear}(1,128) \to\operatorname{GELU}() \to\operatorname{Linear}(128,3N). The GELU activation is chosen because of its smoothness [67]. Supplementary Note A provides more details on the path representation, and Algorithm 2 (Supplementary Note G.2) shows how positions, velocities, energies and forces are evaluated along the path.

Path Initialization

Popcornn implements a continuous path version of the geodesic approach for initial path construction [45]. The geodesic approach uses the pairwise repulsive potential

Ugeo=∑i<jexp⁡[−αrij−rij(eq)rij(eq)]+βrij(eq)rij,(6)U_{\mathrm{geo}} = \sum_{i<j} \exp\left[-\alpha\frac{r_{ij}-r_{ij}^{(\mathrm{eq})}}{r_{ij}^{(\mathrm{eq})}}\right] + \beta\frac{r_{ij}^{(\mathrm{eq})}}{r_{ij}}, \tag*{(6)}

where rijr_{ij} and rij(eq)r_{ij}^{(\mathrm{eq})} are the current and approximate equilibrium distances, respectively, between atoms ii and jj within a cutoff distance of 3 Å, while α=1.7\alpha= 1.7 and β=0.01\beta= 0.01 scale the contributions of each energy term, as originally proposed by Zhu et al. [45]. In the original geodesic approach, discrete points were optimized. In Popcornn, we apply the same loss to a continuous path

Lgeo=∫01∑q∈{(i,j)∣i<j}∣Fq(xθ(t))⋅vθ(t)∣2 dt=∫01∥F(pairs)(xθ(t))⋅vθ(t)∥ dt,(7)\mathcal{L}_{\mathrm{geo}} = \int_{0}^{1} \sqrt{\sum_{q \in\{(i,j)\mid i<j\}} \left|\mathbf{F}_{q}(\mathbf{x}_{\theta}(t)) \cdot\mathbf{v}_{\theta}(t)\right|^{2}}\,\mathrm{d}t = \int_{0}^{1} \left\|\mathbf{F}^{(\mathrm{pairs})}(\mathbf{x}_{\theta}(t)) \cdot\mathbf{v}_{\theta}(t)\right\|\,\mathrm{d}t, \tag*{(7)}

where ∥⋅∥\|\cdot\| is the norm of a (N(N−1)/2)(N(N-1)/2)-dimensional vector across unique atomic pairs, and F(pairs)∈R3N×N(N−1)2\mathbf{F}^{(\mathrm{pairs})} \in\mathbb{R}^{3N \times\frac{N(N-1)}{2}} is a matrix by stacking pairwise forces Fq∈R3N\mathbf{F}_{q} \in\mathbb{R}^{3N}. See Supplementary Note B.4 for more detail. This initial path construction is the first optimization stage of Algorithm 1 (Supplementary Note G.1).

Interfacing with MLIPs

We have interfaced Popcornn with various MLIPs, including NewtonNet [13, 49], MACE [52, 53], ANI-1x [60, 61], LeftNet [51], Orb [62], and UMA [63]. The model weights are acquired from Yuan et al. [13], Zhao et al. [50], and Wood et al. [63]. The Popcornn repository includes a template class and necessary functions to interface with future MLIPs. The Popcornn code repository [47] provides more details, and Supplementary Note G.6 lists where each step is implemented.

TS Determination

When searching for the TS along xθ(t)x_{\theta}(t), we found that both minimizing the PES gradient and maximizing energy led to the most robust method. The TS is the highest energy point along xθ(t)x_{\theta}(t), but simply selecting the highest energy point among the evaluated xθ(t)x_{\theta}(t) points may not be sufficiently precise if the adaptive integration evaluations are too sparse. Instead, we find points between the evaluated path points at which dE/dt=0\mathrm{d}E/\mathrm{d}t = 0 (calculated using Eq. (37)) since this is more stable and has higher precision than a finite difference comparison between energy calculations. First, we determine where dE/dt\mathrm{d}E/\mathrm{d}t changes sign within the cached values from the latest Popcornn optimization evaluation. This can result in multiple TS candidates. Second, we find the TS time of each pair by linearly interpolating between the cached time points and solve for tt that satisfies dE(t)/dt=0\mathrm{d}E(t)/\mathrm{d}t = 0. Finally, we evaluate the PES energy of all the TS time candidates and determine t∗t^* as the highest energy point; this discards sign changes at energy minima. Algorithm 5 (Supplementary Note G.5) gives the full procedure.

Popcornn Path Optimization

Parallel Evaluations and Memory Management

The loss integration is handled by PadaQuad [40]: a parallelized adaptive quadrature numerical integrator for path integrals with backpropagation capabilities. The adaptive numerical integration is necessary to properly evaluate Eq. (1). PadaQuad leverages the knowledge of the path to parallelize the integration, improving speed by orders of magnitude. PadaQuad also uses an adaptive memory management scheme to fully utilize GPU parallelism. We apply the absolute and relative tolerances of 0 and 1, respectively, for the numerical integration. Hegazy et al. [40] and Supplementary Note E provide more detailed discussions. Algorithms 3 and 4 (Supplementary Notes G.3 and G.4) give the gradient evaluation and the adaptive quadrature.

Optimization Hyperparameters

We use the Adam optimizer with a learning rate of 10−310^{-3} for both initial path construction atop the initialization potential and final optimization atop the MLIP potential. Each of these optimizations used its own optimizer; that is, the optimizer was not shared between them. We run each optimization for up to 1000 steps, or stop as it reaches the gradient convergence of 10−210^{-2} for initial path construction and 10−110^{-1} for path optimization (corresponding to TS forces of 0.05 eV/Å). Supplementary Note C provides more detailed discussions, and Algorithm 1 shows where each of these settings enters.

TS Optimization

The TS acquired from Popcornn or other reference methods are optimized to a first-order saddle point on the MLIP using Sella [68] with the full Hessian at every step [13] with a convergence threshold of 0.01 eV/A˚\AA for up to 1000 steps. For retrieved TSs, forward and backwards IRC calculations are run using SellaIRC [68] with a convergence threshold of 0.01 eV/A˚\AA for up to another 1000 steps. For organic reactions, the endpoints are compared using connectivity graphs created by Open Babel [69].

Reference Methods

For a fair comparison, we provide the same inputs to each method and apply the same TS optimization to all outputs. Specifically, we give each method the same optimized endpoints. The output TSs of each method are processed in exactly the same way.

We use “improved tangent” NEB [27] with 10, 20, and 100 beads (k=5.0k=5.0) and GSM [29] with 9 and 25 beads and a convergence threshold of 0.05 eV/A˚\AA. NEBs are optimized with the MDMin optimizer (dt=0.02dt=0.02, maxstep=0.1) in ASE for 2000 steps in total, where the first 1000 steps are without the climbing image and the last 1000 steps are with the climbing image, with a convergence threshold of 0.05 eV/A˚\AA [28]. All other hyperparameters are in ASE 3.28.0 and pyGSM defaults.

Data availability

All reaction data and model weights used in this study are publicly available. The reactions come from the Transition1x dataset [23, 70] and the Grambow–Green dataset [48, 71]; the two test sets are subsets of these datasets, selected as described in Results. The NewtonNet weights trained on Transition1x by Yuan et al. [13] are available at https://github.com/THGLab/NewtonNet/releases/download/pretrained/newtonnet_t1x.zip. The LeftNet and MACE weights trained or fine-tuned on Transition1x by Zhao et al. [50] are available at https://huggingface.co/yhong55/ReactBench, with usage instructions at https://github.com/deepprinciple/ReactBench.

Code availability

Popcornn is available at https://github.com/khegazy/popcornn [47], and the GPU-parallelized adaptive numerical integrator PadaQuad is available at https://github.com/khegazy/padaquad [40]. Both are released under the Creative Commons Attribution 4.0 International licence. The MLIP implementations are available from their public repositories: NewtonNet [49] at https://github.com/THGLab/NewtonNet, LeftNet [51] at https://github.com/yuanqidu/LEFTNet, and MACE [52, 53] at https://github.com/ACEsuit/mace.

Acknowledgements

We thank Graeme Henkelman and Ilgar Baghishov for helpful discussions about NEB hyperparameters. We thank Paul Zimmerman for helpful discussions about GSM hyperparameters. This work used computational resources provided by the National Energy Research Scientific Computing Center (NERSC), a U.S. Department of Energy Office of Science User Facility operated under Contract DE-AC02-05CH11231, and the Lawrencium computational cluster resource provided by the IT Division at the Lawrence Berkeley National Laboratory (Supported by the Director, Office of Science, Office of Basic Energy Sciences, of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231). All authors acknowledge support from the Laboratory Directed Research and Development program of Lawrence Berkeley National Laboratory (LBNL), supported by the Office of Science, Office of Basic Energy Sciences (BES), of the U.S. Department of Energy (DOE) under Contract No. DE-AC02-05CH11231. S.M.B. further acknowledges support from the Center for High Precision Patterning Science (CHiPPS), an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences at Lawrence Berkeley National Laboratory under Contract No. DE-AC02-34205CH11231. C.-Y.E.Y. and T.H.G. further acknowledge support from the CPIMS program and funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences at Lawrence Berkeley National Laboratory under Contract No. DE-AC02-05CH11231. D.M. further acknowledges support from the Scientific Discovery through Advanced Computing (SciDAC) program funded by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research at Lawrence Berkeley National Laboratory under Contract No. DE-AC02-05CH11231. M.W.M. acknowledges DARPA, NSF, the DOE Competitive Portfolios grant, and the DOE SciGPT grant.

Competing interests

The authors declare no competing interests.

Neural Path Representation

We parameterize our reaction path as a learned deviation from a linear interpolation between the endpoints. This parameterization was introduced in Eq. (2) and repeated here for clarity

xθ(t)=(1−t)xR+txP⏟linear prior+t(1−t)fθ(t)⏟perturbation.x_{\theta}(t) = \underbrace{(1-t)\mathbf{x}_{R} + t\mathbf{x}_{P}}_{\text{linear prior}} + \underbrace{t(1-t)f_{\theta}(t)}_{\text{perturbation}}.

This parameterization anchors both the endpoints and the path. Our path model (fθ(t)f_{\theta}(t)) is a simple two-layer MLP with GELU activations. While optimizing the Popcornn path representation xθ(t)x_{\theta}(t), we observed that activation functions that are not smooth (e.g., ReLU and SELU) underperform smooth activations like GELU. We use the Cartesian coordinate system to represent both the input and output structures.

Ablation Study of the Endpoint Enforcement

A double-ended TS finding method must fix the two endpoints, as in Figure 1a. This is trivially accomplished for discrete approaches like NEB, simply by not applying any updates to the endpoint beads. To impose the constraint in the continuous case, however, is not as straightforward. Our linear path formulation (Eq. (2)) enforces the endpoint structures as hard constraints (xθ(0)=xRx_{\theta}(0)=x_{R} and xθ(1)=xPx_{\theta}(1)=x_{P}) through a quadratic envelope. To demonstrate the flexibility and robustness of our methodology, we ablate this envelope by removing it and replacing it with other envelopes.

We remove the endpoint-enforcing envelope and enforce the endpoints by subtracting the neural network values. The resulting path formulation

xθ(t)=(1−t)xR+txP⏟linear prior+fθ(t)−(1−t)fθ(0)−tfθ(1)⏟perturbation,(8)x_{\theta}(t) = \underbrace{(1-t)\mathbf{x}_{R} + t\mathbf{x}_{P}}_{\text{linear prior}} + \underbrace{f_{\theta}(t) - (1-t)f_{\theta}(0) - tf_{\theta}(1)}_{\text{perturbation}}, \tag*{(8)}

enforces xθ(0)=xRx_{\theta}(0)=x_{R} and xθ(1)=xPx_{\theta}(1)=x_{P}. This hard constraint enforcement leads to slight optimization instabilities compared to Eq. (2), shown in Figure A1. We presume these instabilities are due to the interdependence between fθ(t)f_{\theta}(t), fθ(0)f_{\theta}(0), and fθ(1)f_{\theta}(1).

Figure A1. Comparison between path constraint implementations. The path with a quadratic envelope t(1−t)t(1-t) in Eq. (2) (top left) can be optimized more stably than the path with endpoint subtraction in Eq. (8) (top right). The bottom row compares the TS force during dynamics. The colored lines and circles are corresponding measurements at the same time.

We ablate the envelope with functions of varying polynomial degrees and function types. Our ablation envelopes,

Quadratic Envelope:t(1−t)(9)\text{Quadratic Envelope:}\quad t(1-t) \tag*{(9)}
Normalized Quadratic Envelope:4t(1−t)(10)\text{Normalized Quadratic Envelope:}\quad4t(1-t) \tag*{(10)}
Quartic Envelope:t2(1−t)2(11)\text{Quartic Envelope:}\quad t^{2}(1-t)^{2} \tag*{(11)}
Sine Envelope:sin⁡(πt)(12)\text{Sine Envelope:}\quad\sin(\pi t) \tag*{(12)}
Cosine Envelope:1−cos⁡(2πt),(13)\text{Cosine Envelope:}\quad1-\cos(2\pi t), \tag*{(13)}

illustrated in the top panel of Figure A2, primarily differ at the endpoints. We ablate the quadratic term with the normalized quadratic because for the original quadratic envelope xθ(0.5)=0.25x_{\theta}(0.5)=0.25 and for the normalized xθ(0.5)=1x_{\theta}(0.5)=1, similar to the other envelopes. When testing the path optimization on the Muller–Brown potential, all the envelopes behave similarly (Figure A2). When comparing the quadratic and sine envelopes with the quartic and cosine envelopes, the primary difference is at the endpoints. For the quartic and cosine envelopes, the path is artificially straightened. This may become a problem in some real-world optimizations where either the product or reactant state is fairly unstable. This is because the quadratic and sine envelopes have finite velocities at t=0t = 0 and t=1t = 1, whereas the quartic and cosine envelopes have zero velocity. Therefore, we would expect some envelopes to produce a path that cannot be optimized, like a square root (t(1−t)\sqrt{t(1-t)}), which has infinite velocities at the endpoints.

Figure A2. Comparison between path envelope forms. We compare multiple path envelope ablations (top panel) and their optimization dynamics (middle and lower panels). For rigorous comparisons, we scale all envelope functions to have maxima of 1 at t=0.5t = 0.5, roughly equivalent to scaling the learning rate. The colored lines and circles are corresponding measurements at the same time.

Linear Prior Ablation Study

The linear parameterization helps keep the path from deviating to something unphysical. We refer to this as our linear prior since the linear interpolation biases the initial path towards physicality. This linear prior offers a rough direction of the atom movements, and the neural network is trained to correct the 3N3N-dimensional movements as a function of the reaction coordinate tt. This formalism is also adopted by other related work [37, 38, 39].

To illustrate the linear prior’s importance, we consider the bias-variance tradeoff by removing it. The randomized path,

xθ(t)=[1−4t(1−t)][(1−t)xR+txP]+4t(1−t)fθ(t),(14)x_\theta(t) = [1 - 4t(1 - t)][(1 - t)x_R + tx_P] + 4t(1 - t)f_\theta(t), \tag*{(14)}

allows the middle point to be randomly initialized: (xθ(0.5)=fθ(0.5)x_\theta(0.5) = f_\theta(0.5)). This random initialization removes the linear bias, or anchoring effect, of the linear prior. However, this random initialization is likely to initialize xθ(0.5)=fθ(0.5)x_\theta(0.5) = f_\theta(0.5) in a challenging region of the optimization landscape: e.g., at an unphysical structure, a very unlikely structure, or a structure that requires highly unlikely bond breaking and formation. To validate this intuition, we used the random initialization of Eq. (14) on the Muller-Brown potential, where we observe challenging and chaotic optimization (Figure A3). Ultimately, the sporadic optimization of Eq. (14) results in a more convoluted path. Given these challenges with this simple model potential, we expect these issues to be exacerbated when using an MLIP.

Figure A3. Comparison between path formalisms. We show the path optimization atop the model Muller–Brown potential using Popcornn with (top left) and without (top right) the linear prior in Eq. (2). The bottom panel shows the TS force during optimization. The colored lines and circles are corresponding measurements at the same time.

Hidden Units as a Time Basis

Since the action of the xθ(t)x_\theta(t) neural network is a linear combination of the nonlinear hidden units (h(t)h(t)), xθ(t)x_\theta(t) can be interpreted as a weighted sum of learned reaction bases. Here we provide more detail on the path model with our two-layer neural network.

A two-layer neural network, with activations ϕ\phi, has 4 learnable objects: a weight matrix W\mathbf{W} and a bias b\mathbf{b} for each layer. To improve stability for the path neural network, the scalar reaction coordinate tt is first rescaled to

t~=2t−1.(15)\tilde{t} = 2t - 1. \tag*{(15)}

Passing t~\tilde{t} into the first layer of the path network gives us the first set of hidden units

zj(t~)=Wj(1)t~+bj(1),hj(t~)=ϕ(zj(t~)),(16)z_j(\tilde{t}) = W_j^{(1)}\tilde{t} + b_j^{(1)}, \qquad h_j(\tilde{t}) = \phi(z_j(\tilde{t})), \tag*{(16)}

where j=1,…,wj = 1,\ldots,w indexes the hidden units and ϕ\phi is a smooth activation function. The output of the neural network is then

fθ,i(t~)=bi(2)+∑j=1wWij(2)hj(t~),(17)f_{\theta,i}(\tilde{t}) = b_i^{(2)} + \sum_{j=1}^{w} W_{ij}^{(2)}h_j(\tilde{t}), \tag*{(17)}

for Cartesian coordinate ii.

Thus, the hidden units from the first layer (h(t~)h(\tilde{t})) form a trainable basis set of the reaction coordinate. When using a non-smooth activation function (e.g., ReLU or SELU), the basis itself will not be smooth. This additionally makes the atomic velocities and the loss integrand discontinuous, ultimately making the integration, optimization, and convergence more challenging. In our setting with GELU activation, each hj(t~)h_j(\tilde{t}) is a smooth feature over time, and the output layer constructs each Cartesian coordinate of the path as a linear combination of these features.

To enforce the xθ(t)x_\theta(t) endpoints as a hard constraint, we use the quadratic envelope q(t)q(t) (Eq. (9)). Thus, the full xθ(t)x_\theta(t) is

q(t)=t(1−t)(18)q(t) = t(1-t) \tag*{(18)}
xθ,i(t)=(1−t)xR,i+txP,i+q(t)[bi(2)+∑j=1wWij(2)hj(t~)](19)x_{\theta,i}(t) = (1-t)x_{R,i} + tx_{P,i} + q(t)\left[b_i^{(2)} + \sum_{j=1}^{w} W_{ij}^{(2)}h_j(\tilde{t})\right] \tag*{(19)}

This is analogous to representing the path with a set of smooth basis functions in tt, except that the basis functions are learned jointly with the path. Increasing the hidden width increases the number of available time features, while the output layer selects their amplitudes independently for each Cartesian coordinate.

Internal Versus Cartesian Coordinates

Our path operates in Cartesian coordinates instead of the internal coordinates. Although empirical evidence has shown that internal coordinates are important for geometry optimization [68], transforming from internal coordinates to Cartesian is non-unique and generally difficult. This difficulty is in part because internal coordinates can have singularities, and optimal coordinates generally change throughout the reaction.

One of the primary benefits of internal coordinates is the grouped movement of atoms by one or a few coordinates. This feature can be captured by the neural network path, which can learn the optimal time-dependent internal representation. Thus, the neural network path effectively learns the time-dependent internal coordinates and their transformation.

The input molecules must be properly positioned before being passed to Popcornn. This includes handling the translational and rotational degrees of freedom. This is a non-trivial task that is best handled by specialty external packages [64, 72].

Complexity of the Path Model

The neural network fθ(t)f_\theta(t) is much smaller than the MLIP, serving only as a flexible representation of the path, not as a sophisticated regressor. Here, we use a simple fully-connected neural network instead of any invariant or equivariant architecture. Such inductive biases are commonplace in many MLIPs [39, 52]. We found that many of these inductive biases can be detrimental to our path optimization. Moreover, allowing the neural network to learn optimal coordinate systems and transformations removes the need for such inductive biases.

The locality assumption is a very common inductive bias found in most all MLIPs. Here, the representation of each atom is updated by information of atoms within some radius. This, however, can be detrimental for Popcornn, which can optimize the reaction between two separated molecules. Imagine the simplest possible reaction H+H→H2\mathrm{H}+\mathrm{H}\rightarrow\mathrm{H}_2 where the reactant atoms are separated much further than a typical radius used in the locality assumption. This path, therefore, is generally not local. Although some of the non-locality is captured by the linear prior, it is likely that this inductive bias would hamper the optimization.

Another common and important inductive bias is rotation and translation invariance. However, the linear prior breaks the individual translational and rotational symmetries of endpoints. Given the importance of the linear prior in this work, we consider that incorporating invariant or equivariant architectures for the perturbation neural network will lead to complications.

Path Jacobian

The path Jacobian quantifies the sensitivity of each atomic coordinate to the learnable weights, and thus it plays an important role in the optimization. Since the linear prior is independent of θ\theta (Eq. (2)), the path Jacobian is

Jθ(t)=∂xθ(t)∂θ=q(t)∂fθ(t~)∂θ.(20)J_\theta(t)=\frac{\partial\mathbf{x}_\theta(t)}{\partial\theta}=q(t)\frac{\partial f_\theta(\tilde{t})}{\partial\theta}. \tag*{(20)}

The quadratic envelope (Eq. (9)) scales the sensitivity of the path to the neural-network parameters, and removes fθ(t~)f_\theta(\tilde{t})’s contribution from the endpoints. For the two-layer MLP used here,

z(t~)=W(1)t~+b(1),h(t~)=ϕ(z(t~)),(21)\mathbf{z}(\tilde{t})=\mathbf{W}^{(1)}\tilde{t}+\mathbf{b}^{(1)},\qquad\mathbf{h}(\tilde{t})=\phi(\mathbf{z}(\tilde{t})), \tag*{(21)}
fθ(t~)=W(2)h(t~)+b(2),(22)f_\theta(\tilde{t})=\mathbf{W}^{(2)}\mathbf{h}(\tilde{t})+\mathbf{b}^{(2)}, \tag*{(22)}

where ϕ\phi is the GELU activation. If ii indexes Cartesian output coordinates and jj indexes hidden units, the nonzero Jacobian blocks are

∂xθ,i(t)∂Wij(2)=q(t)hj(t~),(23)\frac{\partial x_{\theta,i}(t)}{\partial W^{(2)}_{ij}}=q(t)h_j(\tilde{t}), \tag*{(23)}
∂xθ,i(t)∂bi(2)=q(t),(24)\frac{\partial x_{\theta,i}(t)}{\partial b^{(2)}_i}=q(t), \tag*{(24)}
∂xθ,i(t)∂W(1)=q(t)Wij(2)ϕ′(zj(t~))t~(25)\frac{\partial x_{\theta,i}(t)}{\partial W^{(1)}}=q(t)W^{(2)}_{ij}\phi'(z_j(\tilde{t}))\tilde{t} \tag*{(25)}
∂xθ,i(t)∂bj(1)=q(t)Wij(2)ϕ′(zj(t~)).(26)\frac{\partial x_{\theta,i}(t)}{\partial b^{(1)}_j}=q(t)W^{(2)}_{ij}\phi'(z_j(\tilde{t})). \tag*{(26)}

The final-layer weight block (Eq. (23)) gives a simple estimate of the path Jacobian scale, because the scale of the weights (Wij(2)W^{(2)}_{ij}) is dependent on its input dimension O(1/w)\mathcal{O}(1/\sqrt{w}), while the GELU hidden units hjh_j are O(1)\mathcal{O}(1). Generally, w>100w > 100, making Eq. (23) at least an order of magnitude larger than Eq. (25).

Jθ(t)≈JW(2)(t)=∂xθ,i(t)∂Wij(2).(27)J_\theta(t) \approx J_{W^{(2)}}(t)=\frac{\partial x_{\theta,i}(t)}{\partial W^{(2)}_{ij}}. \tag*{(27)}

Taking the inner product of the Jacobian with itself results in a positive semidefinite matrix, allowing us to make a comparison of the singular values σ\sigma:

σi(Jθ(t)Jθ(t)T)≳σi(JW(2)(t)JW(2)(t)T)=q(t)2∥h(t~)∥22(28)\sigma_i\left(J_\theta(t)J_\theta(t)^\mathsf{T}\right)\gtrsim\sigma_i\left(J_{W^{(2)}}(t)J_{W^{(2)}}(t)^\mathsf{T}\right)=q(t)^2\lVert h(\tilde{t})\rVert_2^2 \tag*{(28)}
σmin⁡(Jθ(t))≳σmin⁡(JW(2)(t))=q(t)∥h(t~)∥2,(29)\sigma_{\min}\left(J_\theta(t)\right)\gtrsim\sigma_{\min}\left(J_{W^{(2)}}(t)\right)=q(t)\lVert h(\tilde{t})\rVert_2, \tag*{(29)}

where the equality holds because the inner product of the Jacobians is diagonal across Cartesian output coordinates. At the path midpoint, q(1/2)=1/4q(1/2)=1/4, t~=0\tilde{t}=0, and h(1/2)=ϕ(b(1))\mathbf{h}(1/2)=\phi(\mathbf{b}^{(1)}). The hidden pre-activations reduce to the first-layer biases:

σmin⁡(Jθ(1/2))≳14∥ϕ(b(1))∥2.(30)\sigma_{\min}\left(J_\theta(1/2)\right)\gtrsim\frac{1}{4}\lVert\phi(\mathbf{b}^{(1)})\rVert_2. \tag*{(30)}

Thus, at the reaction time midpoint, the scale of the smallest singular value is set by both the hidden width and the activation function. For PyTorch’s default initialization of a first-layer map from one scalar input to ww hidden units, the biases satisfy approximately bj(1)∼U[−1,1]b_j^{(1)}\sim\mathcal{U}[-1,1]. Thus,

E[∥h(1/2)∥22]=wEb∼U[−1,1][ϕ(b)2],(31)\mathbb{E}\left[\lVert\mathbf{h}(1/2)\rVert_2^2\right]=w\mathbb{E}_{b\sim\mathcal{U}[-1,1]}\left[\phi(b)^2\right], \tag*{(31)}

and the expected midpoint scale is

σmin⁡(Jθ(1/2))≈14wEb∼U[−1,1][ϕ(b)2].(32)\sigma_{\min}\left(J_\theta(1/2)\right)\approx\frac{1}{4}\sqrt{w\mathbb{E}_{b\sim\mathcal{U}[-1,1]}\left[\phi(b)^2\right]}. \tag*{(32)}

For the GELU activation,

ϕ(b)=b2[1+erf⁡(b2)],(33)\phi(b) = \frac{b}{2}\left[1+\operatorname{erf}\left(\frac{b}{\sqrt{2}}\right)\right], \tag*{(33)}

which gives

Eb∼U[−1,1][ϕ(b)2]=12∫−11ϕ(b)2 db≈0.1088.(34)\mathbb{E}_{b\sim U[-1,1]}\left[\phi(b)^2\right] = \frac{1}{2}\int_{-1}^{1}\phi(b)^2\,\mathrm{d}b \approx0.1088. \tag*{(34)}

Therefore, for the default width w=128w=128,

σmin⁡(Jθ(1/2))≈14128×0.1088≈0.93.(35)\sigma_{\min}(J_\theta(1/2)) \approx\frac{1}{4}\sqrt{128\times0.1088}\approx0.93. \tag*{(35)}

Therefore, both the size of the number of hidden unit bases and the activation itself determine the Jacobian’s scale. As we will show in Supplementary Note C, this Jacobian plays a crucial role in the convergence criteria.

Popcornn Optimization Losses

For double-ended TS searches, we use the assumption of simple TS theory that the path is simply a tool for finding the TS, and other points along the entire path (e.g., MEP, the gradient descent path from the TS, etc.) are not needed. Hence Popcornn is designed to work with all losses that can be written as a path integral (Eq. (1)). In this work, we investigate three path integral losses: LVRE\mathcal{L}_{\mathrm{VRE}} in Eq. (3); LpVRE\mathcal{L}_{\mathrm{pVRE}} in Eq. 4; and Lgeo\mathcal{L}_{\mathrm{geo}} in Eq. 7. We have demonstrated their behaviors in Figure 2, and we provide further discussion on their physical interpretation below.

The Variational Reaction Energy (VRE) Loss

The LVRE\mathcal{L}_{\mathrm{VRE}} loss is only guaranteed to find the gradient descent path from any stationary point to which it converges. This is because the LVRE\mathcal{L}_{\mathrm{VRE}} path can converge onto a high-order saddle point or even a maximum, instead of the TS. This issue is more pronounced on a high-dimensional PES with many stationary points of various orders. To demonstrate this failure mode on a low-dimensional system, we show that the LVRE\mathcal{L}_{\mathrm{VRE}} path converges to the PES maximum of the Wolfe potential, in Figure B4.

Figure B4. Comparison between the LVRE\mathcal{L}_{\mathrm{VRE}} and LpVRE\mathcal{L}_{\mathrm{pVRE}} losses on the Wolfe potential. Optimizing against LVRE\mathcal{L}_{\mathrm{VRE}} (top left) converges the path incorrectly to the gradient descent path from the maximum on the PES, while using LpVRE\mathcal{L}_{\mathrm{pVRE}} (top right) converges the path to going through the TS. The bottom row shows the TS force during optimization. The colored lines and circles are corresponding measurements at the same time.

The Projected Variational Reaction Energy (pVRE) Loss

The projected variational reaction energy (pVRE) loss is most affected by the highest energy point, thus targeting the TS, and better aligns with the assumption of TS theory. The pVRE loss, reproduced here, is

LpVRE=∫01∣F(xθ(t))⋅vθ(t)∣ dt=∑a2Ea,TS−Ea,R−Ea,P,(36)\mathcal{L}_{\mathrm{pVRE}}=\int_{0}^{1}\left|\mathbf{F}(\mathbf{x}_\theta(t))\cdot\mathbf{v}_\theta(t)\right|\,\mathrm{d}t=\sum_a 2E_{a,\mathrm{TS}}-E_{a,\mathrm{R}}-E_{a,\mathrm{P}}, \tag*{(36)}

where aa sums over the maxima along the path, ERE_{\mathrm{R}} is the corresponding reactant energy, and EPE_{\mathrm{P}} is the corresponding product energy. LpVRE\mathcal{L}_{\mathrm{pVRE}} can be understood as the total variation of the energy profile, i.e., the summation of energy differences along xθ(t)x_{\theta}(t). Define the energy profile along the path as E(xθ(t))E(x_{\theta}(t)). Since

dE(xθ(t))dt=dE(xθ(t))dxθ(t)⋅dxθ(t)dt=−F(xθ(t))⋅vθ(t),(37)\frac{dE(x_{\theta}(t))}{dt}=\frac{dE(x_{\theta}(t))}{dx_{\theta}(t)}\cdot\frac{dx_{\theta}(t)}{dt}=-\mathbf{F}(x_{\theta}(t))\cdot\mathbf{v}_{\theta}(t), \tag*{(37)}

the pVRE loss can be written as the total variation of the energy profile

LpVRE=∫01∣dE(xθ(t))dt∣ dt.(38)\mathcal{L}_{\mathrm{pVRE}}=\int_{0}^{1}\left|\frac{dE(x_{\theta}(t))}{dt}\right|\,dt. \tag*{(38)}

A LpVRE\mathcal{L}_{\mathrm{pVRE}} optimized path will prefer a single TS, generating a smooth path that avoids substructures and multi-step reactions that can be challenging for optimization methods to navigate. We demonstrate this in the Muller-Brown potential (Figure 2) and the Wolfe potential (Figure B4), where xθ(t)x_{\theta}(t) targets the highest TS and avoids the others. This behavior is valuable since the largest barrier (TS) corresponds to the reaction’s rate-determining step, even for a multi-step reaction. If one is interested in the elementary reactions, they can still be retrieved through IRCs from the TS [12].

This preference for transition-state regions can be seen directly from the loss gradient:

∂∂θLpVRE=∫01sgn⁡(dE(xθ(t))dt)∂∂θ(dE(xθ(t))dt) dt.(39)\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}} = \int_{0}^{1} \operatorname{sgn}\left(\frac{dE(x_\theta(t))}{dt}\right) \frac{\partial}{\partial\theta}\left(\frac{dE(x_\theta(t))}{dt}\right) \,dt. \tag*{(39)}

Given that the path endpoints are fixed,

∂E(xθ(0))∂θ=∂E(xθ(1))∂θ=0,(40)\frac{\partial E(x_\theta(0))}{\partial\theta} = \frac{\partial E(x_\theta(1))}{\partial\theta} = 0, \tag*{(40)}

and integrating by parts gives

∂∂θLpVRE=−∫01[ddtsgn⁡(dE(xθ(t))dt)]∂E(xθ(t))∂θ dt.(41)\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}} = -\int_{0}^{1}\left[\frac{d}{dt}\operatorname{sgn}\left(\frac{dE(x_\theta(t))}{dt}\right)\right]\frac{\partial E(x_\theta(t))}{\partial\theta}\,dt. \tag*{(41)}

The derivative of the sign function is a Dirac delta distribution,

ddtsgn⁡(dE(xθ(t))dt)=2d2E(xθ(t))dt2δD(dE(xθ(t))dt),(42)\frac{d}{dt}\operatorname{sgn}\left(\frac{dE(x_\theta(t))}{dt}\right) = 2\frac{d^{2}E(x_\theta(t))}{dt^{2}}\delta_{\mathrm{D}}\left(\frac{dE(x_\theta(t))}{dt}\right), \tag*{(42)}

where δD\delta_{\mathrm{D}} denotes the Dirac delta. Therefore, the gradient is nonzero only where

dE(xθ(t))dt=0,(43)\frac{dE(x_\theta(t))}{dt}=0, \tag*{(43)}

i.e., at stationary points of the energy profile along xθ(t)x_\theta(t). For stationary points tat_a, this can be written as

∂∂θLpVRE=−2∑asgn⁡(d2E(xθ(ta))dt2)∂E(xθ(ta))∂θ.(44)\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}} = -2\sum_{a}\operatorname{sgn}\left(\frac{d^{2}E(x_\theta(t_a))}{dt^{2}}\right)\frac{\partial E(x_\theta(t_a))}{\partial\theta}. \tag*{(44)}

If xθ(t)x_\theta(t) has only one energy barrier, the sum would only have one concave stationary point, which is the proposed TS, and can be simplified as

∂∂θLpVRE=2∂E(xθ(t∗))∂θ,(45)\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}} = 2\frac{\partial E(x_\theta(t^*))}{\partial\theta}, \tag*{(45)}

where 0<t∗<10 < t^* < 1 is the proposed TS time along xθ(t)x_\theta(t). Using

∂E(xθ(t))∂θ=Jθ(t)T∂E(xθ(t))∂xθ(t)=−Jθ(t)TF(xθ(t)),(46)\frac{\partial E(x_\theta(t))}{\partial\theta} = J_\theta(t)^{\mathsf{T}}\frac{\partial E(x_\theta(t))}{\partial x_\theta(t)} = -J_\theta(t)^{\mathsf{T}}\mathbf{F}(x_\theta(t)), \tag*{(46)}

where Jθ(t)=∂xθ(t)/∂θJ_\theta(t) = \partial x_\theta(t)/\partial\theta defined in A13, we obtain

∂∂θLpVRE=−2Jθ(t∗)TF(xθ(t∗)).(47)\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}} = -2J_\theta(t^*)^{\mathsf{T}}\mathbf{F}(x_\theta(t^*)). \tag*{(47)}

Therefore, LpVRE\mathcal{L}_{\mathrm{pVRE}} removes all maxima and minima along xθ(t)x_\theta(t), and the TS at t∗t^* is the only contributing time point to the loss. For interior maxima, the sign of Eq. (47) lowers the PES energy. For interior minima, the sign of Eq. (47) raises the PES energy. Therefore, xθ(t)x_\theta(t) is optimized to avoid any interior maxima or minima. However, the TS cannot be removed, and thus it becomes the only contributing extremum to minimizing LpVRE\mathcal{L}_{\mathrm{pVRE}}. This provides a complementary view of why pVRE focuses optimization on transition-state structure: flat or monotonic portions of the path do not contribute to the first-order signal, whereas energy extrema contribute through delta-function weights.

This behavior is further intuited by the fact that xθ(t)x_\theta(t) accumulates no loss when traveling on an isoenergy line. For xθ(t)x_\theta(t) to traverse an isoenergy line, its velocity must be perpendicular to the force, and by Eq. 4 the integrand is 0. The ability for xθ(t)x_\theta(t) to freely travel along isoenergy paths aligns with the statistical nature of chemical reactions, where structures of the same energy are equally likely to occur and thus xθ(t)x_\theta(t) should be able to freely travel between them.

A Combined Variational Loss: Variational Error

The work of Birkholz and Schlegel [36] defines a non-negative variational error loss to find the steepest descent reaction path. The variational error

εVRE=LVRE−LpVRE(48)\varepsilon_{\mathrm{VRE}} = \mathcal{L}_{\mathrm{VRE}} - \mathcal{L}_{\mathrm{pVRE}} \tag*{(48)}

is the difference between the VRE and pVRE losses. Minimizing this loss extremizes the dot product of LpVRE\mathcal{L}_{\mathrm{pVRE}}: aligning the gradient of the PES with the velocity. However, all gradient descent paths minimize εVRE\varepsilon_{\mathrm{VRE}}; in fact, Eq. (48) is independent of the stationary point. That is, one can reach εVRE=0\varepsilon_{\mathrm{VRE}} = 0 by going over any feature where the PES gradient is 0: any PES peak, any saddle point, any PES minima. Therefore, once xθ(t)x_\theta(t) corresponds to one of these extrema, there is likely no way to optimize xθ(t)x_\theta(t) towards the correct TS. Instead, the path will be optimized to align xθ(t)x_\theta(t) with the PES gradient while staying at this extreme.

Losses for Path Initialization: Geodesic Loss

Properly initializing the reaction path is a crucial step (Figure 3b), and doing so incorrectly can place xθ(t)x_\theta(t) in a PES region where it will be optimized to the incorrect TS. That is, the path initialization effectively confines the optimization to a PES region and must therefore be sensitive to the region with the correct TS. This is necessary for multiple reasons: often accurate optimizations can be more expensive and cannot optimize over large PES regions; MLIPs can become undefined in some PES regions; and it is generally prohibitively challenging or expensive to optimize xθ(t)x_\theta(t) across many large PES structures to a new PES region on complex PES surfaces using especially ab initio methods. Therefore, a quality path initializer must leverage chemical biases to select the PES region, balance PES evaluation cost, and often smooth the PES to allow easier optimization. Hence, one often uses a pairwise potential (Eq. (6)) that is fast to evaluate, smooths the PES in a principled way, and still maintains chemical expectations. Optimizing the initial path over the pairwise PES with LpVRE\mathcal{L}_{\mathrm{pVRE}} and LVRE\mathcal{L}_{\mathrm{VRE}} suffers from catastrophic failures caused by approximations in the pairwise PES. For LpVRE\mathcal{L}_{\mathrm{pVRE}} this stems from the fact that on the pairwise PES the reactants and products are generally not at the minima, and there is often no TS in between. Since Eq. 4 has a monotonic integrand, the integral would be a constant E(xθ(1))−E(xθ(0))E(x_{\theta}(1))-E(x_{\theta}(0)) regardless of the path. Thus LpVRE\mathcal{L}_{\mathrm{pVRE}} is insensitive to the initial path in the common scenarios where reactant, product, and TS stationary points are missing and therefore cannot optimize the path. In contrast, the LVRE\mathcal{L}_{\mathrm{VRE}} path is sensitive to the entire path, instead of a few special points like LpVRE\mathcal{L}_{\mathrm{pVRE}}. However, the total energy is not very indicative since the lengthening of one bond will be canceled by the shortening of another bond for a pairwise potential like Eq. (6). This makes LVRE\mathcal{L}_{\mathrm{VRE}} insensitive to initial path changes. Additionally, the energy can be reduced by compensating one bond motion with the opposite motion from another bond; this increases the possibility of visiting an unlikely structure where DFT and MLIPs may become undefined.

The geodesic loss (Lgeo\mathcal{L}_{\mathrm{geo}}) [45] can be intuited as minimizing the amount of displacement atoms must make while being rearranged from the reactant to the product structure. This imposes a chemical intuition that the optimal reaction path can generally be optimized from the “least-motion” or the spatially shortest path. Instead of summing the pairwise energy contributions before evaluating the force or the loss integrand, Lgeo\mathcal{L}_{\mathrm{geo}} squares these individual pairwise contributions to minimize them separately

Lgeo=∫01∑q∈{(i,j)∣i<j}∣Fq(xθ(t))⋅vθ(t)∣2 dt=∫01∥F(pairs)(xθ(t))⋅vθ(t)∥ dt,(49)\mathcal{L}_{\mathrm{geo}} = \int_0^1 \sqrt{\sum_{q \in\{(i,j)\mid i<j\}} \left|\mathbf{F}_q(x_\theta(t)) \cdot\mathbf{v}_\theta(t)\right|^2}\,dt = \int_0^1 \left\|\mathbf{F}^{(\mathrm{pairs})}(x_\theta(t)) \cdot\mathbf{v}_\theta(t)\right\|\,dt, \tag*{(49)}

Figure B5 demonstrates the difference between paths optimized by Lgeo\mathcal{L}_{\mathrm{geo}} and LpVRE\mathcal{L}_{\mathrm{pVRE}} on a 3-dimensional example. The path minimizing Lgeo\mathcal{L}_{\mathrm{geo}} does not correspond to the path crossing the minimal barrier minimizing LpVRE\mathcal{L}_{\mathrm{pVRE}}.

Figure B5. Comparison between the geodesic path and the minimum energy path. Let a hypothetical PES has three energy contributions ExE_x, EyE_y, and EzE_z, and the feasible manifold is a sphere. The path with the smallest barrier, which is barrier-less in this case, lies on the plane x+y+z=1x+y+z=1 and follows the energy contour. The geodesic, i.e., the path with the shortest arc, lies on the plane y=0y=0.

To demonstrate the correspondence between Lgeo\mathcal{L}_{\mathrm{geo}} and LpVRE\mathcal{L}_{\mathrm{pVRE}}, we derive Lgeo\mathcal{L}_{\mathrm{geo}} from LpVRE\mathcal{L}_{\mathrm{pVRE}}. To see this, consider a generalized pVRE loss as a sum of the pVRE losses from each pairwise energy term,

Lgeo(1)=∑q∈{(i,j)∣i<j}LpVRE,q(50)\mathcal{L}_{\mathrm{geo}}^{(1)} = \sum_{q \in\{(i,j)\mid i<j\}} \mathcal{L}_{\mathrm{pVRE},q} \tag*{(50)}
=∑q∈{(i,j)∣i<j}∫01∣Fq(xθ(t))⋅vθ(t)∣ dt(51)= \sum_{q \in\{(i,j)\mid i<j\}} \int_0^1 \left|\mathbf{F}_q(x_\theta(t)) \cdot\mathbf{v}_\theta(t)\right|\,dt \tag*{(51)}
=∫01∑q∈{(i,j)∣i<j}∣Fq(xθ(t))⋅vθ(t)∣ dt(52)= \int_0^1 \sum_{q \in\{(i,j)\mid i<j\}} \left|\mathbf{F}_q(x_\theta(t)) \cdot\mathbf{v}_\theta(t)\right|\,dt \tag*{(52)}
=∫01∥F(pairs)(xθ(t))⋅vθ(t)∥1 dt,(53)= \int_0^1 \left\|\mathbf{F}^{(\mathrm{pairs})}(x_\theta(t)) \cdot\mathbf{v}_\theta(t)\right\|_1\,dt, \tag*{(53)}

where ∥⋅∥1\lVert\cdot\rVert_1 is the 1-norm of a (N(N−1)/2)(N(N-1)/2)-dimensional vector, and F(pairs)∈R3N×N(N−1)2\mathbf{F}^{(\mathrm{pairs})} \in\mathbb{R}^{3N \times\frac{N(N-1)}{2}} is a matrix by stacking pairwise forces Fq∈R3N\mathbf{F}_q \in\mathbb{R}^{3N}, instead of summing them so that we can treat all pairs independently. Physically speaking, if one bond remains unchanged from the reactant to the product, this should be reflected in the learned path, and the contribution to the integral in Eq. (53) should be zero. This Lgeo(1)\mathcal{L}_{\mathrm{geo}}^{(1)}, as a generalized LpVRE\mathcal{L}_{\mathrm{pVRE}}, can still be independent of the path on the pairwise potential energy surface. To add a path dependence, we can replace the 1-norm with the 2-norm, i.e., the Euclidean norm, in a sense that one replaces the displacement with the distance or replaces the velocity with the speed,

Lgeo(2)=∫01∑q∈{(i,j)∣i<j}∣Fq(xθ(t))⋅vθ(t)∣2 dt(54)\mathcal{L}_{\mathrm{geo}}^{(2)}=\int_0^1\sqrt{\sum_{q\in\{(i,j)\mid i<j\}}\left\lvert\mathbf{F}_q(\mathbf{x}_{\theta}(t))\cdot\mathbf{v}_{\theta}(t)\right\rvert^2}\,\mathrm{d}t \tag*{(54)}
=∫01∥F(pairs)(xθ(t))⋅vθ(t)∥2 dt,(55)=\int_0^1\left\lVert\mathbf{F}^{(\mathrm{pairs})}(\mathbf{x}_{\theta}(t))\cdot\mathbf{v}_{\theta}(t)\right\rVert_2\,\mathrm{d}t, \tag*{(55)}

which leads us to the geodesic loss Lgeo\mathcal{L}_{\mathrm{geo}} in Eq. 7, to find the shortest path from the reactant to the product in this N(N−1)/2N(N-1)/2-dimensional space of energy terms. We note that Hait et al. [73] have applied the geodesic formalism to acquire the pVRE loss, without deeper physical explanation, but we find the opposite direction of understanding from LpVRE\mathcal{L}_{\mathrm{pVRE}} to Lgeo\mathcal{L}_{\mathrm{geo}} is more physically intuitive.

Transition State Losses

Popcornn is equipped with TS losses to speed up path converge [36] and improve upon the TS structure. Currently, these losses fall into two categories: the TS loss (LTS\mathcal{L}_{\mathrm{TS}}) and the TS region loss (LTSR\mathcal{L}_{\mathrm{TSR}})

LTS=ℓTS(tTS)(56)\mathcal{L}_{\mathrm{TS}} = \ell_{\mathrm{TS}}(t_{\mathrm{TS}}) \tag*{(56)}
LTSR=∫titfℓTS(t) dt,(57)\mathcal{L}_{\mathrm{TSR}} = \int_{t_i}^{t_f} \ell_{\mathrm{TS}}(t)\,\mathrm{d}t, \tag*{(57)}

where ℓTS\ell_{\mathrm{TS}} is any user-defined function (including those used to optimize xθ(t)x_{\theta}(t)) applicable to the TS or TS region, tTSt_{\mathrm{TS}} is the TS time, and ti<tTS<tft_i < t_{\mathrm{TS}} < t_f defines a region of the path around the TS. More obvious choices for ℓTS\ell_{\mathrm{TS}} that are supported by Popcornn include the potential energy and the magnitude of its gradient at the TS. The latter loss helps optimize the path to align with the TS saddle point.

The LTSR\mathcal{L}_{\mathrm{TSR}} can be practically useful during path optimization by lowering the energy or the PES gradient of the highest region in the path. Birkholz and Schlegel [36] showed that using a TS loss during optimization helped the path converge much faster. The LTSR\mathcal{L}_{\mathrm{TSR}} also helps define the TS when using LpVRE\mathcal{L}_{\mathrm{pVRE}}. Specifically, when applying LpVRE\mathcal{L}_{\mathrm{pVRE}} optimization, xθ(t)x_{\theta}(t) is not penalized for traversing isoenergy geometries. Therefore, many structures near the TS can have the same or similar energies and potentially confuse a TS searching scheme. For example, selecting the TS based on energy can be challenging, as the deviation from the maximum energy may be small in the TS region. By minimizing the energy in the TS region, the TS will become more distinguishable.

In this work, we ran tests with both LTS\mathcal{L}_{\mathrm{TS}} and LTSR\mathcal{L}_{\mathrm{TSR}}, but ultimately we did not use them. While the TS losses speed up the reaction path convergence [36], they lead to inferior paths due to faster convergence. Many of these paths corresponded to multi-step reactions. Given our goal is to find single-step reaction TS, the TS losses were not well suited for our specific task. We do believe that the TS losses can be well-suited for other tasks, where one is interested in multi-step reactions or in higher TS structure precision. We also note that the TS losses can be turned on at or near the end of the Popcornn optimization (Supplementary Note D) to avoid path optimization effects while still improving the TS itself.

Popcornn Convergence

For each optimization stage, Popcornn monitors the path-integrated gradient of the loss with respect to the path parameters (Eq. (47)). A stage is declared converged when

∥∂∂θLpVRE∥2<τ(58)\left\lVert\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}} \right\rVert_2 < \tau \tag*{(58)}

for pp consecutive optimization steps, where τ\tau is the convergence threshold and pp is the patience parameter. If this condition is not met, the stage terminates after the prescribed maximum number of optimizer iterations.

The threshold τ\tau is chosen to correspond to a target residual force at the predicted TS. The residual is anything that is nonzero because TS should have zero forces, while in practice a small nonzero residual is commonly allowed. Let

F2target(59)F_2^{\mathrm{target}} \tag*{(59)}

denote the desired force magnitude of ∥F(x(t∗))∥2\lVert\mathbf{F}(x(t^*))\rVert_2 at the TS candidate on the optimized path. We wish to stop the optimization according to Eq. (47) at

∥∂∂θLpVRE∥2=2∥Jθ(t∗)TF(xθ(t∗))∥2<τ(60)\left\lVert\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}}\right\rVert_2 = 2\left\lVert\mathbf{J}_{\theta}(t^*)^{\mathsf{T}}\mathbf{F}(x_{\theta}(t^*))\right\rVert_2 < \tau \tag*{(60)}

Using the singular-value bound given by

∥Jθ(t∗)TF(xθ(t∗))∥22=F(xθ(t∗))TJθ(t∗)Jθ(t∗)TF(xθ(t∗))≥σmin⁡(Jθ(t∗))2∥F(xθ(t∗))∥22(61)\begin{aligned} \left\lVert\mathbf{J}_{\theta}(t^*)^{\mathsf{T}}\mathbf{F}(x_{\theta}(t^*))\right\rVert_2^2 &= \mathbf{F}(x_{\theta}(t^*))^{\mathsf{T}}\mathbf{J}_{\theta}(t^*)\mathbf{J}_{\theta}(t^*)^{\mathsf{T}}\mathbf{F}(x_{\theta}(t^*)) \\ &\ge\sigma_{\min}(\mathbf{J}_{\theta}(t^*))^2\left\lVert\mathbf{F}(x_{\theta}(t^*))\right\rVert_2^2 \tag*{(61)} \end{aligned}

we have

∥Jθ(t∗)TF(xθ(t∗))∥2≥σmin⁡(Jθ(t∗))∥F(xθ(t∗))∥2.(62)\left\lVert\mathbf{J}_{\theta}(t^*)^{\mathsf{T}}\mathbf{F}(x_{\theta}(t^*))\right\rVert_2 \ge\sigma_{\min}(\mathbf{J}_{\theta}(t^*))\left\lVert\mathbf{F}(x_{\theta}(t^*))\right\rVert_2. \tag*{(62)}

For LpVRE\mathcal{L}_{\mathrm{pVRE}}, this bound yields the estimate

τ>∥∂∂θLpVRE∥2≥2σmin⁡(Jθ(t∗))∥F(xθ(t∗))∥2(63)\tau> \left\lVert\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}}\right\rVert_2 \ge2\sigma_{\min}(\mathbf{J}_{\theta}(t^*))\left\lVert\mathbf{F}(x_{\theta}(t^*))\right\rVert_2 \tag*{(63)}

Thus, we bound the loss gradient using the singular value lower bound and our force target F2targetF_2^{\mathrm{target}}

τ:=2σmin⁡(Jθ(t∗))F2target.(64)\tau:= 2\sigma_{\min}(\mathbf{J}_{\theta}(t^*))F_2^{\mathrm{target}}. \tag*{(64)}

With the common F2targetF_2^{\mathrm{target}} of 0.05 eV/Ang and the singular value σmin⁡(Jθ(t∗))≈1\sigma_{\min}(\mathbf{J}_{\theta}(t^*)) \approx1 is estimated by A28, we set τ=0.1\tau= 0.1 throughout this work.

While xθ(t)x_{\theta}(t) converges, the integration tolerance must be at most the same magnitude as τ\tau. At convergence, the optimization should reach a steady state at the optimal xθ(t)x_{\theta}(t) and the gradient of the loss with respect to the gradients is 0. As previously mentioned, we bound this by τ\tau, and therefore at convergence

∥∂∂θLpVRE∥2≈τ(65)\left\lVert\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}}\right\rVert_2 \approx\tau \tag*{(65)}

Since the adaptive quadrature tolerance is given by

max⁡(atol,rtol∥∂∂θLpVRE∥2),(66)\max\left(a_{\mathrm{tol}},r_{\mathrm{tol}}\left\lVert\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}}\right\rVert_2\right), \tag*{(66)}

we set atol=0a_{\mathrm{tol}} = 0 so the integration convergence is dependent on the loss objective and its gradient. This determines rtolr_{\mathrm{tol}} as

rtol∥∂∂θLpVRE∥2≈τrtol=1.(67)r_{\mathrm{tol}}\left\lVert\frac{\partial}{\partial\theta}\mathcal{L}_{\mathrm{pVRE}}\right\rVert_{2}\approx\tau\quad r_{\mathrm{tol}}=1. \tag*{(67)}

The integration tolerance is therefore matched to the stopping threshold: small enough not to mask convergence, but loose enough to avoid unnecessary quadrature refinement during convergence.

Curriculum Optimization and Popcornn GSM

Popcornn can apply dynamic weighting to the path integral loss in a curriculum learning approach. Curriculum learning is a machine learning technique that dynamically alters the loss to emphasize specific contributions the user wishes the optimization to focus on [74, 75]. Heuristically, this is similar to the growing string method (GSM), which optimizes endpoints of strings that are continuously growing from the initial and final points. Popcornn is equipped with a growing-string-like curriculum that initially increases/decreases the path loss contributions near to and far from the endpoints. The curriculum ends by evenly weighting all path points.

Popcornn is also capable of increasing or decreasing specific loss contributions at different points of the optimization. For example, the LTS\mathcal{L}_{\mathrm{TS}} can be absent in the first 3/43/4 of the path optimization (avoiding convergence to an inferior path) and then increased at the end to improve the TS. Similarly, paths optimized with LpVRE\mathcal{L}_{\mathrm{pVRE}} converge faster and find better TS than LVRE\mathcal{L}_{\mathrm{VRE}}, but can ignore PES structures by traversing isoenergy lines. This can make a LpVRE\mathcal{L}_{\mathrm{pVRE}} path difficult to interpret and will not resolve any local minima. To leverage the superior LpVRE\mathcal{L}_{\mathrm{pVRE}} performance and create a more interpretable path that resolves local minima (intermediate equilibrium structures), one can add LVRE\mathcal{L}_{\mathrm{VRE}} to the LpVRE\mathcal{L}_{\mathrm{pVRE}} loss near the end of the optimization.

Parallelized Adaptive Numerical Integration and PadaQuad

Adaptive numerical integration is necessary to leverage the continuous path xθ(t)\mathbf{x}_{\theta}(t). When xθ(t)\mathbf{x}_{\theta}(t) is evaluated on a fixed grid of reaction times, the loss depends only on the structures at the grid points, which act as discrete images. These structures are set by both the shape of the path and its velocity profile, i.e., how quickly the path moves through each region. Changing the velocity profile, “re-timing” the path, moves the images along the path without changing its shape. By changing the velocity profile, images “fall” toward regions of low loss during optimization. For LE\mathcal{L}_{\mathrm{E}}, the images fall down the PES toward low-energy structures such as xR\mathbf{x}_{\mathrm{R}} and xP\mathbf{x}_{\mathrm{P}}. NEB prevents this by adding fictitious spring forces that keep the images evenly spaced [27, 28], effectively fixing the velocity profile [38].

For LVRE\mathcal{L}_{\mathrm{VRE}} and LpVRE\mathcal{L}_{\mathrm{pVRE}}, re-timing alone can localize the images at stationary points of the PES, such as xR\mathbf{x}_{\mathrm{R}}, xP\mathbf{x}_{\mathrm{P}}, and the TS, where the force vanishes. Although the exact LVRE\mathcal{L}_{\mathrm{VRE}} and LpVRE\mathcal{L}_{\mathrm{pVRE}} integrals are independent of the velocity profile, their fixed-grid sums are not because re-timing changes which structures sit at the grid points. The optimizer can therefore lower the fixed-grid loss without improving the path by placing the grid points at stationary points while the path crosses the high-loss regions between them. In the limit, xθ(t)\mathbf{x}_{\theta}(t) rests at xR\mathbf{x}_{\mathrm{R}} and xP\mathbf{x}_{\mathrm{P}} for nearly the entire reaction time and crosses between them within a single grid interval. The images then coincide at the endpoints, and the fixed-grid loss vanishes regardless of the barrier the path crosses.

Instead of fixing the velocity profile through a fictitious spring force, the adaptive integrator adds evaluation points where the integration error is large and removes them where it is small, until the error falls below a set tolerance. Where the path moves quickly, the integrand generally varies rapidly; the integrator adds points in these places, so re-timing no longer lowers the loss. For LVRE\mathcal{L}_{\mathrm{VRE}} and LpVRE\mathcal{L}_{\mathrm{pVRE}}, adaptive integration therefore removes the need for spring forces and keeps the loss within a set tolerance of the exact integral, so the gradient depends on the shape of the path rather than its velocity profile.

PadaQuad

Standard numerical integrators used for machine learning tasks are generally designed for ODEs [76] and require sequential evaluations in time, which would make Popcornn computationally intractable. Instead, we use PadaQuad [40]: an adaptive numerical quadrature method designed to parallelize path integration on GPUs with backpropagation compatibility. PadaQuad, as an adaptive numerical quadrature method, divides the integral into steps that are evaluated and refined in parallel. Steps are refined depending on their error (εstep\varepsilon_{\mathrm{step}}). When given a new integral, PadaQuad will simultaneously evaluate integration steps based on a randomly sampled grid, shown as black points in Figure 1b. Regions with an error larger than the tolerance will be split in half, and new evaluation points will be made and evaluated in parallel (gray points in Figure 1b). This procedure continues until all integration steps satisfy the error criteria (white points in Figure 1b) or have been split a maximum number of times (5 by default; Algorithm 4). The error criteria

εstep<Max⁡(atol,rtol×L′),(68)\varepsilon_{\mathrm{step}} < \operatorname{Max}(\mathrm{atol}, \mathrm{rtol} \times\mathcal{L}^\prime), \tag*{(68)}

where atol\mathrm{atol} and rtol\mathrm{rtol} are the absolute and relative tolerances, and L′\mathcal{L}^\prime is the latest path integral value. The atol\mathrm{atol} and rtol\mathrm{rtol} values determine the integrator’s accuracy. To evaluate an integration step’s error from Eq. (68), the integral (L′\mathcal{L}^\prime) must be known. PadaQuad defines L′\mathcal{L}^\prime as the integral with the current integration steps. As the integration steps are refined, L′\mathcal{L}^\prime is updated with each parallel evaluation. Thus, the integral value and the grid refinement converge simultaneously.

To further increase speed when integrating the same function, PadaQuad uses the previous evaluation points as an initial evaluation for the next optimization step. PadaQuad avoids evaluating excess temporal points by pruning the evaluation time points at the end of each integration. Note that PadaQuad uses all evaluated integration steps in the returned integral value; this pruned grid is only used to initialize the integration steps in the next optimization. When initializing the integral with the previous integration times, PadaQuad generally requires 2 or 3 adaptive evaluations in the systems we tested.

Popcornn is designed to handle systems that require large memory budgets, which may stem from expensive potentials or from a many-atom system. To do this, PadaQuad calculates the maximum number of evaluations that can fit onto the GPU, and as a redundancy, reduces the evaluation batch size on out-of-memory errors. PadaQuad can reduce the batch size to a single evaluation, thus handling any system where a single evaluation can fit within the GPU’s memory. The user may also set the evaluation batch size themselves to optimize Popcornn on their own hardware.

Hyperparameter Screening

Setting up the reference NEB and GSM methods is, in general, difficult and often case dependent. We aim to make the comparison as fair as possible under a fixed set of hyperparameters. Below are some of the most important discretization and initialization hyperparameters we sampled. All others are kept at default in ASE 3.28.0.

NEB setting screening on Grambow-Green out-of-distribution reactions

Figure F6. NEB setting screening on Grambow-Green out-of-distribution reactions. We evaluate NEB’s performance with different hyperparameters. We explore the number of beads: (a) 10 beads, (b) 20 beads, and (c) 100 beads. For each number of beads, we vary the initial path construction method (linear interpolation, IDPP interpolation, discretized geodesic interpolation, and Popcornn-geodesic). The number at the bottom of each bar is the average time, in seconds, for the TS optimization to converge.

Popcornn Algorithms and Pseudocode

This Supplementary Note gives pseudocode for Popcornn as implemented in the code repository [47]; we checked it against release 0.2.1 of the code (commit 15dd905, 15 November 2026) and release 1.4.1 of the PadaQuad implementation padaquad. Algorithm 1 is a complete Popcornn run, and Algorithms 2–5 expand its steps. Settings stated in Methods appear here as inputs. Settings not stated in Methods take the code defaults: the patience p=1p = 1, the number of output structures nrec=101n_{\mathrm{rec}} = 101, the random seed 0, single-precision (32-bit) arithmetic, the maximum number of times an integration step may be split Msplit=5M_{\mathrm{split}} = 5, and the error norm ∥⋅∥ν\lVert\cdot\rVert_{\nu} with ν=2\nu= 2 (the Euclidean norm of a vector). Throughout, NN is the number of atoms, θ\theta denotes the parameters of the path network fθ(t)f_{\theta}(t) with DD entries in total, and t∈[0,1]t \in[0,1] is the reaction time. Derivatives are computed by automatic differentiation, which obtains exact derivatives of a computer program by applying the chain rule to each of its operations; reverse-mode automatic differentiation (backpropagation) gives the derivative of a scalar with respect to all of θ\theta in one pass.

A Complete Popcornn Run

A Popcornn run is a sequence of optimization stages that share the path parameters θ\theta. For the molecular benchmarks in this work there are two stages, initial path construction and path optimization on the MLIP (Methods), each with its own potential, integrand, and Adam optimizer. Every optimizer iteration integrates the gradient of the path loss along the whole path (Algorithm 3) and updates all of θ\theta at once. After the last iteration of a stage, one more integration records dE/dtdE/dt along the path, from which the TS is located (Algorithm 5); the TS that Popcornn reports is the one found after the final stage. Optionally, the loss integral and the TS search are also computed at every iteration for monitoring; they do not change θ\theta.

Evaluating the Path

Algorithm 2 evaluates the path model of Eq. (2) at a batch of reaction times, with the network input rescaled to t~=2t−1\tilde{t} = 2t - 1 (Supplementary Note A.3). All quantities are computed for the whole batch at once and stay differentiable with respect to θ\theta. The velocity vθ(t)v_{\theta}(t) is the derivative of the path with respect to tt, obtained by automatic differentiation rather than by finite differences. The potential is called once per batch; the repository provides one wrapper per supported MLIP that returns energies and atomic forces for a batch of structures. For the force-based losses, the gradient with respect to θ\theta passes through the forces, so it needs the derivative of the forces with respect to the atomic positions. For models whose forces are the negative energy gradient, this derivative is the negative of the PES Hessian (the matrix of second derivatives of the energy), and automatic differentiation applies it to vectors without forming the full matrix. For models that predict forces directly, or for the analytic forces of the pair potential, automatic differentiation differentiates that force expression instead.

Gradient of the Path Loss

Algorithm 3 computes the parameter gradient by integrating the gradient of the integrand, ∫01∂ℓ/∂θ dt\int_{0}^{1} \partial\ell/\partial\theta\, dt in Eq. (5), as described in Results; this is the code default, and the other option, differentiating the completed integral of ℓ\ell, is not used in this work. The Algorithm 1 A complete Popcornn run.

Require: reactant and product structures xR,xP∈R3Nx_R, x_P \in\mathbb{R}^{3N}, relaxed and aligned (Methods); optional intermediate structures x1,…,xKx_1,\ldots,x_K; stages s=1,…,Ss=1,\ldots,S, each with a potential UsU_s, an integrand ℓs\ell_s, absolute and relative quadrature tolerances atol⁡s\operatorname{atol}_s and rtol⁡s\operatorname{rtol}_s, a learning rate ηs\eta_s, an optional convergence threshold τs\tau_s, a patience psp_s, and a maximum number of iterations MsM_s

Ensure: the path at nrecn_{\mathrm{rec}} evenly spaced times, the TS estimate xθ(t∗)x_\theta(t^*), and for each stage whether its convergence test was met

  1. Set the random seed; initialize θ\theta with PyTorch’s default initialization for linear layers

  2. if intermediate structures are given then ▷ not used in this work

  3. tk←k/(K+1)t_k \leftarrow k/(K+1) for k=0,…,K+1k=0,\ldots,K+1, with x0=xRx_0=x_R and xK+1=xPx_{K+1}=x_P

  4. Fit θ\theta with Adam (learning rate 10−310^{-3}) to minimize 13N(K+2)∑k=0K+1∥xθ(tk)−xk∥2\frac{1}{3N(K+2)}\sum_{k=0}^{K+1}\lVert x_\theta(t_k)-x_k\rVert^2, stopping when it falls below 10−810^{-8} or after 10410^4 iterations

  5. end if

  6. for s=1,…,Ss=1,\ldots,S do

  7. Attach UsU_s to the path; create a new Adam optimizer with learning rate ηs\eta_s; c←0c\leftarrow0 ▷ cc: consecutive iterations below τs\tau_s

  8. for m=1,…,Msm=1,\ldots,M_s do

  9. G←PathGradient⁡(θ,Us,ℓs,atol⁡s,rtol⁡s)G \leftarrow\operatorname{PathGradient}(\theta,U_s,\ell_s,\operatorname{atol}_s,\operatorname{rtol}_s) ▷ Algorithm 3; G≈∂Ls/∂θG\approx\partial\mathcal{L}_s/\partial\theta

  10. if τs\tau_s is set and ∥G∥ν<τs\lVert G\rVert_\nu<\tau_s then

  11. c←c+1c\leftarrow c+1

  12. else

  13. c←0c\leftarrow0

  14. end if

  15. if c≥psc\geq p_s then

  16. break ▷ converged (Supplementary Note C)

  17. end if

  18. θ←AdamStep⁡(θ,G,ηs)\theta\leftarrow\operatorname{AdamStep}(\theta,G,\eta_s) ▷ an optional learning-rate schedule also steps here

  19. end for

  20. (G,Q)←PathGradient⁡(θ,Us,ℓs,atol⁡s,rtol⁡s)(G,Q)\leftarrow\operatorname{PathGradient}(\theta,U_s,\ell_s,\operatorname{atol}_s,\operatorname{rtol}_s), recording the samples QQ of dE/dt\mathrm{d}E/\mathrm{d}t (Algorithm 3); also integrate the loss for reporting, with its own tolerances (atol⁡=10−4\operatorname{atol}=10^{-4}, rtol⁡=0\operatorname{rtol}=0)

  21. t∗←TSSearch⁡(Q)t^*\leftarrow\operatorname{TSSearch}(Q) ▷ Algorithm 5

  22. end for

  23. return xθ(ti)x_\theta(t_i) for ti=(i−1)/(nrec−1)t_i=(i-1)/(n_{\mathrm{rec}}-1), i=1,…,nreci=1,\ldots,n_{\mathrm{rec}}, and xθ(t∗)x_\theta(t^*)

integrand passed to the adaptive quadrature is therefore the vector ∂ℓ(t)/∂θ∈RD\partial\ell(t)/\partial\theta\in\mathbb{R}^D. For a batch of nn times, the nn gradient vectors come from one batched reverse-mode pass. When requested, the same evaluations also record dE/dt=−F⋅v\mathrm{d}E/\mathrm{d}t=-\mathbf{F}\cdot\mathbf{v} (Eq. (37)) at every evaluated time. Only the samples in integration steps that the quadrature finally accepts are kept; samples in steps that it splits are discarded. Algorithm 5 uses the kept samples.

Algorithm 2 Evaluating the path at a batch of reaction times.

  1. function EvaluatePath(t1,…,tn;θ,Ut_1,\ldots,t_n;\theta,U)

  2. for all j=1,…,nj=1,\ldots,n in parallel do

  3. xj←(1−tj)xR+tjxP+tj(1−tj)fθ(2tj−1)\mathbf{x}_j \leftarrow(1-t_j)\mathbf{x}_R+t_j\mathbf{x}_P+t_j(1-t_j)f_\theta(2t_j-1) ▷ Eq. (2)

  4. Set the network perturbation of atoms fixed by constraints to zero

  5. vj←dxj/dt\mathbf{v}_j \leftarrow d\mathbf{x}_j/dt, by automatic differentiation

  6. For periodic cells, wrap xj\mathbf{x}_j into the cell before calling the potential ▷ not used in this work

  7. (Ej,Fj)←U(xj)(E_j,\mathbf{F}_j)\leftarrow U(\mathbf{x}_j) ▷ Fj=−∇Ej\mathbf{F}_j=-\nabla E_j for conservative models

  8. For the pair potential (Eq. (6)), also return the pairwise forces Fj(pairs)∈R3N×N(N−1)/2\mathbf{F}_j^{(\mathrm{pairs})}\in\mathbb{R}^{3N\times N(N-1)/2}

  9. end for

  10. return those of {xj,vj,Ej,Fj,Fj(pairs)}j=1n\{\mathbf{x}_j,\mathbf{v}_j,E_j,\mathbf{F}_j,\mathbf{F}_j^{(\mathrm{pairs})}\}_{j=1}^n that the caller asks for

  11. end function

Algorithm 3 Gradient of the path loss with respect to the path parameters.

  1. function PathGradient(θ,U,ℓ,atol,rtol\theta,U,\ell,a_{\mathrm{tol}},r_{\mathrm{tol}})

  2. function g(t1,…,tn)g(t_1,\ldots,t_n) ▷ integrand: one gradient vector per time

  3. Evaluate the path at t1,…,tnt_1,\ldots,t_n ▷ Algorithm 2

  4. ℓj←∑mwmℓm,j\ell_j\leftarrow\sum_m w_m\ell_{m,j} over the configured terms; this work uses one term with w=1w=1: ∣Fj⋅vj∣|\mathbf{F}_j\cdot\mathbf{v}_j| (pVRE, Eq. 4), ∥Fj∥∥vj∥\|\mathbf{F}_j\|\|\mathbf{v}_j\| (VRE, Eq. (3)), or ∥(Fj(pairs))⊤vj∥\|(\mathbf{F}_j^{(\mathrm{pairs})})^\top\mathbf{v}_j\| (geodesic, Eq. 7)

  5. Record tjt_j, −Fj⋅vj-\mathbf{F}_j\cdot\mathbf{v}_j in QQ ▷ only when samples are requested

  6. gj←∂ℓj/∂θ\mathbf{g}_j\leftarrow\partial\ell_j/\partial\theta for all jj, by batched reverse-mode automatic differentiation

  7. return [g1,…,gn]∈Rn×D[\mathbf{g}_1,\ldots,\mathbf{g}_n]\in\mathbb{R}^{n\times D}

  8. end function

  9. G←G\leftarrow AdaptiveQuadrature(g,atol,rtolg,a_{\mathrm{tol}},r_{\mathrm{tol}}) ▷ Algorithm 4

  10. Keep in QQ only the samples in accepted integration steps, and sort them by time

  11. return GG (and QQ)

  12. end function

Adaptive Quadrature

Algorithm 4 gives the parallel adaptive quadrature of PadaQuad as described in Supplementary Note E and [40]. Popcornn calls it through the padaquad package, an implementation of PadaQuad that is not part of the Popcornn repository; this algorithm follows that source. The integration interval is divided into integration steps. Each step carries an integral estimate and an error estimate from an embedded rule pair: two quadrature rules of different order that share evaluation points, whose difference estimates the error. Popcornn’s default, used in this work, is the 7-point Gauss–Kronrod rule with its embedded 3-point Gauss rule. The integrand is evaluated at 9 points per step: the 7 Gauss–Kronrod points and the step’s two ends, which have zero weight in both rules. Because the integrand is the gradient vector, the integral

Algorithm 4 Parallel adaptive quadrature (PadaQuad).

1: function AdaptiveQuadrature(gg, atola_{\mathrm{tol}}, rtolr_{\mathrm{tol}}) 2: if this is the first call of the stage then 3: New ←\leftarrow 11 integration steps: [0,1][0,1] is cut into 3 equal segments, and each segment is cut at 3 uniformly random times 4: B←B \leftarrow the user’s batch-size cap if set, else, before each round of evaluations, the largest number of evaluations that fits in the free GPU memory, predicted by a memory model fitted at the start of the call 5: else ▷ warm start 6: New ←\leftarrow the steps stored by the previous call 7: B←B \leftarrow the batch size stored by the previous call 8: end if 9: Accepted ←∅\leftarrow\varnothing 10: repeat 11: Evaluate gg at the 9 points of every step in New, in batches of at most BB points ▷ if GPU memory runs out, retry with B←max⁡(1,min⁡(round⁡(0.75B),B−1))B \leftarrow\max(1,\min(\operatorname{round}(0.75B),B-1)) 12: For each step kk in New, compute its integral IkI_k and error εk\varepsilon_k from the embedded rule pair 13: G′←∑k∈Accepted∪NewIkG' \leftarrow\sum_{k\in\mathrm{Accepted}\cup\mathrm{New}} I_k ▷ current estimate of the integral 14: Failed ←{k∈Accepted∪New:∥εk∥ν≥max⁡(atol,rtol∥G′∥ν,50ϵ∥Ik∥ν)\leftarrow\{k\in\mathrm{Accepted}\cup\mathrm{New}:\lVert\varepsilon_k\rVert_\nu\ge\max(a_{\mathrm{tol}},r_{\mathrm{tol}}\lVert G'\rVert_\nu,50\epsilon\lVert I_k\rVert_\nu) and step kk has been split fewer than MsplitM_{\mathrm{split}} times}$ ▷ Eq. (68); ϵ\epsilon: machine epsilon of the working precision 15: Accepted ←\leftarrow (Accepted ∪\cup New) ∖\setminus Failed 16: New ←\leftarrow the two halves of every failed step 17: until New is empty 18: Store BB and the accepted steps for the next call, after merging neighbouring pairs whose merged error, predicted from the integrand values already computed, is below 10% of the tolerance and halving the steps that still fail because they reached MsplitM_{\mathrm{split}} 19: return G′←∑k∈AcceptedIkG' \leftarrow\sum_{k\in\mathrm{Accepted}} I_k 20: end function

estimate G′G' (the L′\mathcal{L}' of Eq. (68)) and the step errors are vectors in RD\mathbb{R}^{D}, and the error criterion compares their ν\nu-norms; the same norm is used for the convergence test in Algorithm 1. With atol=0a_{\mathrm{tol}}=0, as in this work, the criterion depends only on rtolr_{\mathrm{tol}}, except that a step is never required to be more accurate than floating-point rounding allows. Steps accepted earlier are tested again against each new G′G', because G′G' changes as refinement proceeds. A step that has already been split MsplitM_{\mathrm{split}} times is accepted even if it fails the criterion. After the loop, every accepted step is tested once more against the final G′G'; if any step fails, padaquad logs a warning and marks the integral as not converged, but Popcornn uses the integral unchanged. An integrand value that is not finite stops the run with an error. Each Popcornn stage creates a new integrator, so warm starting (Supplementary Note E) and a reduced batch size carry over between the calls of a stage but not between stages.

Locating the Transition State

Algorithm [5] is the full TS search summarized in Methods. It uses the samples recorded during the final integration of a stage and brackets each zero of dE/dtdE/dt between neighbouring samples. Because each integration step is evaluated at its two ends, the first and last samples lie exactly at the reactant and product (t=0t = 0 and t=1t = 1). These two samples are excluded from the brackets because, at and close to the reactant and product minima, dE/dtdE/dt is near zero and its sign is unreliable. The reactant and product times are always candidates, so that a path without an interior maximum returns the higher-energy endpoint. Taking the highest-energy candidate also discards zeros of dE/dtdE/dt that are energy minima.

Algorithm 5 Locating the TS along the optimized path.

Require: samples Q={(ti,di)}i=1nQ = \{(t_i,d_i)\}_{i=1}^{n} sorted by tit_i, with di=dE/dtd_i = dE/dt at tit_i

  1. function TSSearch(QQ)

  2. C←{0,1}\mathcal{C} \leftarrow\{0,1\} ▷ reactant and product times

  3. for i=2,…,n−2i = 2,\ldots,n-2 do

  4. &nbsp;&nbsp;&nbsp;&nbsp;if didi+1≤0d_i d_{i+1} \le0 and not di=di+1=0d_i = d_{i+1} = 0 then ▷ sign change of dE/dtdE/dt

  5. &nbsp;&nbsp;&nbsp;&nbsp;C←C∪{ti−di(ti+1−ti)/(di+1−di)}\mathcal{C} \leftarrow\mathcal{C} \cup\left\{t_i - d_i(t_{i+1}-t_i)/(d_{i+1}-d_i)\right\} ▷ zero of dE/dtdE/dt

  6. &nbsp;&nbsp;&nbsp;&nbsp;end if

  7. end for

  8. Evaluate EE and F\mathbf{F} at every t∈Ct \in\mathcal{C} in one batched call ▷ Algorithm [2]

  9. t∗←arg⁡max⁡t∈CE(xθ(t))t^* \leftarrow\arg\max_{t\in\mathcal{C}} E(x_\theta(t))

  10. return t∗t^*, E(xθ(t∗))E(x_\theta(t^*)), F(xθ(t∗))\mathbf{F}(x_\theta(t^*)), and the barrier E(xθ(t∗))−E(xR)E(x_\theta(t^*)) - E(x_\mathrm{R})

  11. end function

Where Each Step Is Implemented

Table G1 maps each algorithm to the module and function of the Popcornn repository [47] that implement it. A new MLIP is added by writing one wrapper in popcornn/potentials/ that returns energies and forces for a batch of structures, and adding its name to get_potential in popcornn/potentials/__init__.py.

AlgorithmModule (function)
1, input processingpopcornn/tools/images.py (process_images)
1, seed and initial fitpopcornn/popcornn.py (Popcornn.__init__), popcornn/optimization/initialize_path.py (initialize_path)
1, stages and iteration looppopcornn/popcornn.py (Popcornn.optimize_path, optimize)
1, Adam step and convergencepopcornn/optimization/optimizer.py (PathOptimizer)
1, outputspopcornn/popcornn.py (get_discrete_path, get_ts)
2, path modelpopcornn/paths/ (mlp.py, linear.py, base_path.py)
2, potentialspopcornn/potentials/ (pair_potential: repel.py)
3, integrandspopcornn/tools/integrand.py
3, gradient integrationpopcornn/tools/integrator.py (PathIntegrator)
4, adaptive quadraturepadaquad.adaptive_quadrature, called from integrator.py
5, TS searchpopcornn/paths/base_path.py (BasePath.ts_search)

Table G1. Location of each algorithm in the Popcornn repository (release 0.2.1) and in padaquad (release 1.4.1).

References

  1. [1]Johansson, K.O., Head-Gordon, M.P., Schrader, P.E., Wilson, K.R., Michelsen, H.A.: Resonance-stabilized hydrocarbon-radical chain reactions may explain soot inception and growth. Science 361, 997–1000 (2018) https://doi.org/10.1126/science.aat3417
  2. [2]Martí, C., Michelsen, H.A., Najm, H.N., Zádor, J.: Comprehensive kinetics on the c7h7 potential energy surface under combustion conditions. The Journal of Physical Chemistry A 127(8), 1941–1959 (2023) https://doi.org/10.1021/acs.jpca.2c08035
  3. [3]Xie, X., Spotte-Smith, E.W.C., Wen, M., Patel, H.D., Blau, S.M., Persson, K.A.: Data-driven prediction of formation mechanisms of lithium ethylene monocarbonate with an automated reaction network. Journal of the American Chemical Society 143, 13245–13258 (2021) https://doi.org/10.1021/jacs.1c05807
  4. [4]Spotte-Smith, E.W.C., Kam, R.L., Barter, D., Xie, X., Hou, T., Dwaraknath, S., Blau, S.M., Persson, K.A.: Toward a mechanistic model of solid-electrolyte interphase formation and evolution in lithium-ion batteries. ACS Energy Letters 7, 1446–1453 (2022) https://doi.org/10.1021/acsenergylett.2c00517
  5. [5]Barter, D., Spotte-Smith, E.W.C., Redkar, N.S., Khanwale, A., Dwaraknath, S., Persson, K.A., Blau, S.M.: Predictive stochastic analysis of massive filter-based electrochemical reaction networks. Digital Discovery 2, 123–137 (2023) https://doi.org/10.1039/D2DD00117A
  6. [6]Spotte-Smith, E.W.C., Blau, S.M., Barter, D., Leon, N.J., Hahn, N.T., Redkar, N.S., Zavadil, K.R., Liao, C., Persson, K.A.: Chemical reaction networks explain gas evolution mechanisms in mg-ion batteries. Journal of the American Chemical Society 145, 12181–12192 (2023) https://doi.org/10.1021/jacs.3c02222
  7. [7]Spotte-Smith, E.W.C., Petrocelli, T.B., Patel, H.D., Blau, S.M., Persson, K.A.: Elementary decomposition mechanisms of lithium hexafluorophosphate in battery electrolytes and interphases. ACS Energy Letters 8, 347–355 (2023) https://doi.org/10.1021/acsenergylett.2c02351
  8. [8]Garza, A.J., Bell, A.T., Head-Gordon, M.: Mechanism of co2 reduction at copper surfaces: Pathways to c2 products. ACS Catalysis 8, 1490–1499 (2018) https://doi.org/10.1021/accatal.7b03477
  9. [9]Arnold, P.L., Ochiai, T., Lam, F.Y.T., Kelly, R.P., Seymour, M.L., Maron, L.: Metallacyclic actinide catalysts for dinitrogen conversion to ammonia and secondary amines. Nature Chemistry 2020 12:7 12, 654–659 (2020) https://doi.org/10.1038/s41557-020-0457-9
  10. [10]Demarteau, J., Epstein, A.R., Christensen, P.R., Abubekerov, M., Wang, H., Teat, S.J., Seguin, T.J., Chan, C.W., Scown, C.D., Russell, T.P., Keasling, J.D., Persson, K.A., Helms, B.A.: Circularity in mixed-plastic chemical recycling enabled by variable rates of polydiketoenamine hydrolysis. Science Advances 8, 8823 (2022) https://doi.org/10.1126/sciadv.abp8823
  11. [11]Schlegel, H.B.: Exploring potential energy surfaces for chemical reactions: An overview of some practical methods. Journal of Computational Chemistry 24(12), 1514–1527 (2003) https://doi.org/10.1002/jcc.10231
  12. [12]Schlegel, H.B.: Geometry optimization. Wiley Interdisciplinary Reviews: Computational Molecular Science 1, 790–809 (2011) https://doi.org/10.1002/wcms.34
  13. [13]Yuan, E.C.Y., Kumar, A., Guan, X., Hermes, E.D., Rosen, A.S., Zádor, J., Head-Gordon, T., Blau, S.M.: Analytical ab initio hessian from a deep learning potential for transition state optimization. Nature Communications 2024 15:1 15, 1–9 (2024) https://doi.org/10.1038/s41467-024-52481-5
  14. [14]Burger, A., Thiede, L., Rønne, N., Bernales, V., Vijaykumar, N., Vegge, T., Bhowmik, A., Aspuru-Guzik, A.: Shoot from the HIP: Hessian Interatomic Potentials without derivatives (2025). https://arxiv.org/abs/2509.21624
  15. [15]Wander, B., Musielewicz, J., Cheula, R., Kitchin, J.R.: Accessing Numerical Energy Hessians with Graph Neural Network Potentials and Their Application in Heterogeneous Catalysis (2024). https://arxiv.org/abs/2410.01650
  16. [16]Wu, G., Yuan, E.C.Y., Hegazy, K., Blau, S.M., Head-Gordon, T.: Machine-learned leftmost hessian eigenvectors for robust transition state finding. Journal of Chemical Theory and Computation 22(13), 6670–6678 (2026) https://doi.org/10.1021/acs.jctc.6c00554
  17. [17]Vijver, R.V., Zádor, J.: Kinbot: Automated stationary point search on potential energy surfaces. Computer Physics Communications 248, 106947 (2020) https://doi.org/10.1016/j.cpc.2019.106947
  18. [18]Duan, C., Du, Y., Jia, H., Kulik, H.J.: Accurate transition state generation with an object-aware equivariant elementary reaction diffusion model. Nature Comp. Sci. 3(12), 1045–1055 (2023) https://doi.org/10.1038/s43588-023-00563-7
  19. [19]Kim, S., Woo, J., Kim, W.Y.: Diffusion-based generative AI for exploring transition states from 2D molecular graphs. Nature Communications 15(1), 341 (2024) https://doi.org/10.1038/s41467-023-44629-6
  20. [20]Duan, C., Liu, G.-H., Du, Y., Chen, T., Zhao, Q., Jia, H., Gomes, C.P., Theodorou, E.A., Kulik, H.J.: Optimal transport for generating transition states in chemical reactions. Nature Machine Intelligence 7(4), 615–626 (2025) https://doi.org/10.1038/s42256-025-01010-0
  21. [21]Galustian, L., Mark, K., Karwounopoulos, J., Kovar, M.P.-P., Heid, E.: GoFlow: efficient transition state geometry prediction with flow matching and E(3)-equivariant neural networks. Digital Discovery 4(12), 3492–3501 (2025) https://doi.org/10.1039/d5dd00283d
  22. [22]Han, Y., Ren, G., Liu, Y., Zhang, R., Huang, H., Zhang, R., Yu, Z., Liu, C., Li, Z., Lin, X., Wang, J., Qin, Y., Wang, M., Mei, G., Wang, D., Cui, T., Jia, H., Duan, C., Zhao, Q.: Generative discovery of chemical reaction networks (2026). https://arxiv.org/abs/2610.38236
  23. [23]Schreiner, M., Bhowmik, A., Vegge, T., Busk, J., Winther, O.: Transition1x - a dataset for building generalizable reactive machine learning potentials. Scientific Data 2022 9:1 9, 1–9 (2022) https://doi.org/10.1038/s41597-022-01870-w
  24. [24]Blau, S.M., Patel, H.D., Spotte-Smith, E.W.C., Xie, X., Dwaraknath, S., Persson, K.A.: A chemically consistent graph architecture for massive reaction networks applied to solid-electrolyte interphase formation. Chemical Science 12, 4931–4939 (2021) https://doi.org/10.1039/D0SC05647B
  25. [25]Wen, M., Spotte-Smith, E.W.C., Blau, S.M., McDermott, M.J., Krishnapriyan, A.S., Persson, K.A.: Chemical reaction networks and opportunities for machine learning. Nature Computational Science 2023 3:1 3, 12–24 (2023) https://doi.org/10.1038/s43588-022-00369-z
  26. [26]Zhao, Q., Savoie, B.M.: Simultaneously improving reaction coverage and computational cost in automated reaction prediction tasks. Nature Computational Science 1(7), 479–490 (2021) https://doi.org/10.1038/s43588-021-00101-3
  27. [27]Jónsson, H., Mills, G., Jacobsen, K.W.: Nudged elastic band method for finding minimum energy paths of transitions. In: Berne, B.J., Ciccotti, G., Coker, D.F. (eds.) Classical and Quantum Dynamics in Condensed Phase Simulations, pp. 385–404. World Scientific, Singapore (1998). https://doi.org/10.1142/9789812839664_0016
  28. [28]Henkelman, G., Uberuaga, B.P., Jónsson, H.: A climbing image nudged elastic band method for finding saddle points and minimum energy paths. The Journal of Chemical Physics 113, 9901–9904 (2000) https://doi.org/10.1063/1.1329672
  29. [29]Zimmerman, P.M.: Growing string method with interpolation and optimization in internal coordinates: Method and examples. Journal of Chemical Physics 138, 184102 (2013) https://doi.org/10.1063/1.4804162
  30. [30]Zimmerman, P.: Reliable transition state searches integrated with the growing string method. Journal of Chemical Theory and Computation 9, 3043–3050 (2013) https://doi.org/10.1021/ct400319w
  31. [31]Zimmerman, P.M.: Single-ended transition state finding with the growing string method. Journal of Computational Chemistry 36, 601–611 (2015) https://doi.org/10.1002/jcc.23833
  32. [32]Jafari, M., Zimmerman, P.M.: Reliable and efficient reaction path and transition state finding for surface reactions with the growing string method. Journal of Computational Chemistry 38, 645–658 (2017) https://doi.org/10.1002/jcc.24720
  33. [33]Fukui, K.: The path of chemical reactions - the irc approach. Accounts of Chemical Research 14, 363–368 (1981) https://doi.org/10.1021/ar00072a001
  34. [34]Sheppard, D., Terrell, R., Henkelman, G.: Optimization methods for finding minimum energy paths. Journal of Chemical Physics 128 (2008) https://doi.org/10.1063/1.2841941
  35. [35]Goodman, J.M., Silva, M.A.: Qrc: a rapid method for connecting transition structures to reactants in the computational analysis of organic reactivity. Tetrahedron Letters 44, 8233–8236 (2003) https://doi.org/10.1016/j.tetlet.2003.09.074
  36. [36]Birkholz, A.B., Schlegel, H.B.: Path optimization by a variational reaction coordinate method. i. development of formalism and algorithms. J Chem Phys 143(24), 244101 (2015)DOI
  37. [37]Han, J., Wu, Z., Gu, S., Zhou, X.: Stringnet: Neural network based variational method for transition pathways. Communications in Computational Physics 41(1), 116–152 (2027) https://doi.org/10.4208/cicp.OA-2025-0056
  38. [38]Ramakrishnan, K., Schaaf, L.L., Lin, C., Wang, G., Torr, P.: Implicit neural representations for chemical reaction paths. Journal of Chemical Physics 163, 34109 (2025) https://doi.org/10.1063/5.0267023
  39. [39]Nam, J., Steiner, M., Misterka, M., Yang, S., Singhal, A., Gómez-Bombarelli, R.: Transferable learning of reaction pathways from geometric priors. J. Phys. Chem. Lett. 16, 11690–11699 (2025)arxiv.org/abs/2504.15370
  40. [40]Hegazy, K., Morozov, D., Blau, S., Mahoney, M.W.: PadaQuad. GitHub (2025). https://github.com/khegazy/padaquad
  41. [41]Crehuet, R., Bofill, J.M.: The reaction path intrinsic reaction coordinate method and the hamilton-jacobi theory. Journal of Chemical Physics 122, 234105 (2005) https://doi.org/10.1063/1.1927521
  42. [42]Aguilar-Mogas, A., Crehuet, R., Giménez, X., Bofill, J.M.: Applications of analytic and geometry concepts of the theory of calculus of variations to the intrinsic reaction coordinate model. Molecular Physics 105, 2475–2492 (2007) https://doi.org/10.1080/00268970701519762
  43. [43]Quapp, W.: Chemical reaction paths and calculus of variations. Theoretical Chemistry Accounts 121, 227–237 (2008) https://doi.org/10.1007/s00214-008-0468-x
  44. [44]Rong, Z., Kitchaev, D., Canepa, P., Huang, W., Ceder, G.: An efficient algorithm for finding the minimum energy path for cation migration in ionic materials. Journal of Chemical Physics 145 (2016) https://doi.org/10.1063/1.4960790
  45. [45]Zhu, X., Thompson, K.C., Martínez, T.J.: Geodesic interpolation for reaction pathways. Journal of Chemical Physics 150, 164103 (2019) https://doi.org/10.1063/1.5090303
  46. [46]Smidstrup, S., Pedersen, A., Stokbro, K., Jónsson, H.: Improved initial guess for minimum energy path calculations. Journal of Chemical Physics 140, 214106 (2014) https://doi.org/10.1063/1.4878664
  47. [47]Hegazy, K., Yuan, E.C.-Y., Kumar, A., Morozov, D., Mahoney, M., Head-Gordon, T., Blau, S.M.: Popcorn: Path Optimization with a Continuous Representation Neural Network. Version 0.2.1 (2026). https://github.com/khegazy/popcorn
  48. [48]Grambow, C.A., Pattanaik, L., Green, W.H.: Reactants, products, and transition states of elementary chemical reactions based on quantum chemistry. Scientific Data 2020 7:1 7, 1–8 (2020) https://doi.org/10.1038/s41597-020-0460-4
  49. [49]Haghighatlari, M., Li, J., Guan, X., Zhang, O., Das, A., Stein, C.J., Heidar-Zadeh, F., Liu, M., Head-Gordon, M., Bertels, L., Hao, H., Leven, I., Gordon, T.H.: Newtonnet: a newtonian message passing network for deep learning of interatomic potentials and forces. Digital Discovery 1, 333–343 (2022) https://doi.org/10.1039/D2DD00008C
  50. [50]Zhao, Q., Han, Y., Zhang, D., Wang, J., Zhong, P., Cui, T., Yin, B., Cao, Y., Jia, H., Duan, C.: Harnessing machine learning to enhance transition state search with interatomic potentials and generative models. Advanced Science, 06240 (2025) https://doi.org/10.1002/advs.202506240
  51. [51]Du, W., Du, Y., Wang, L., Feng, D., Wang, G., Ji, S., Gomes, C.P., Ma, Z.M.: A new perspective on building efficient and expressive 3d equivariant graph neural networks. Advances in Neural Information Processing Systems 36 (2023)arxiv.org/abs/2304.04757
  52. [52]Nigam, J., Pozdnyakov, S., Fraux, G., Ceriotti, M.: Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Journal of Chemical Physics 156 (2022) https://doi.org/10.1063/5.0087042
  53. [53]Kovács, D.P., Moore, J.H., Browning, N.J., Batatia, I., Horton, J.T., Pu, Y., Kapil, V., Witt, W.C., Magdău, I.-B., Cole, D.J., Csányi, G.: Mace-off: Transferable short range machine learning force fields for organic molecules. Chemical Physics 1, 1–12 (2023)
  54. [54]Queiruga, A.F., Erichson, N.B., Taylor, D., Mahoney, M.W.: Continuous-in-Depth Neural Networks (2020). https://arxiv.org/abs/2008.02389
  55. [55]Krishnapriyan, A.S., Gholami, A., Zhe, S., Kirby, R.M., Mahoney, M.W.: Characterizing possible failure modes in physics-informed neural networks (2021). https://arxiv.org/abs/2109.01050
  56. [56]Krishnapriyan, A.S., Queiruga, A.F., Erichson, N.B., Mahoney, M.W.: Learning continuous models for continuous physics (2023). https://arxiv.org/abs/2202.08494
  57. [57]Sakarvadia, M., Hegazy, K., Totounferoush, A., Chard, K., Yang, Y., Foster, I., Mahoney, M.W.: The False Promise of Zero-Shot Super-Resolution in Machine-Learned Operators (2026). https://arxiv.org/abs/2510.06646
  58. [58]Benitez, J.A.L., Hegazy, K., Guo, J., Dokmanić, I., Mahoney, M.W., Hoop, M.V.: Neural equilibria for long-term prediction of nonlinear conservation laws (2026). https://arxiv.org/abs/2501.06933
  59. [59]Hegazy, K., Mahoney, M.W., Erichson, N.B.: Recency Biased Causal Attention for Time-series Forecasting (2026). https://arxiv.org/abs/2502.06151
  60. [60]Smith, J.S., Isayev, O., Roitberg, A.E.: Ani-1: an extensible neural network potential with dft accuracy at force field computational cost. Chemical Science 8, 3192–3203 (2017) https://doi.org/10.1039/C6SC05720A
  61. [61]Smith, J.S., Nebgen, B., Lubbers, N., Isayev, O., Roitberg, A.E.: Less is more: Sampling chemical space with active learning. Journal of Chemical Physics 148, 241733 (2018) https://doi.org/10.1063/1.5023802
  62. [62]Neumann, M., Gin, J., Rhodes, B., Bennett, S., Li, Z., Choubisa, H., Hussey, A., Godwin, J.: Orb: A Fast, Scalable Neural Network Potential (2024). https://arxiv.org/abs/2410.22570
  63. [63]Wood, B.M., Dzamba, M., Fu, X., Gao, M., Shuaibi, M., Barroso-Luque, L., Abdelmaqsoud, K., Gharakhanyan, V., Kitchin, J.R., Levine, D.S., Michel, K., Sriram, A., Cohen, T., Das, A., Rizvi, A., Sahoo, S.J., Ulissi, Z.W., Zitnick, C.L.: UMA: A Family of Universal Models for Atoms (2026). https://arxiv.org/abs/2506.23971
  64. [64]Unsleber, J.P., Grimmel, S.A., Reiher, M.: Chemoton 2.0: Autonomous exploration of chemical reaction networks. Journal of Chemical Theory and Computation 18, 5393–5409 (2022) https://doi.org/10.1021/acs.jctc.2c00193
  65. [65]Levine, D.S., Shuaibi, M., Spotte-Smith, E.W.C., Taylor, M.G., Hasyim, M.R., Michel, K., Batatia, I., Csányi, G., Dzamba, M., Eastman, P., Frey, N.C., Fu, X., Gharakhanyan, V., Krishnapriyan, A.S., Rackers, J.A., Raja, S., Rizvi, A., Rosen, A.S., Ulissi, Z., Vargas, S., Zitnick, C.L., Blau, S.M., Wood, B.M., Meta, F.: The Open Molecules 2025 (OMol25) dataset, evaluations, and models. arXiv preprint arXiv:2505.08762 (2025)arxiv.org/abs/2505.08762
  66. [66]Yuan, E.C.-Y., Liu, Y., Chen, J., Zhong, P., Raja, S., Kreiman, T., Vargas, S., Xu, W., Head-Gordon, M., Yang, C., Blau, S.M., Cheng, B., Krishnapriyan, A., Head-Gordon, T.: Foundation Models for Atomistic Simulation of Chemistry and Materials (2025)arxiv.org/abs/2503.10538
  67. [67]Hendrycks, D., Gimpel, K.: Gaussian Error Linear Units (GELUs) (2023). https://arxiv.org/abs/1606.08415
  68. [68]Hermes, E.D., Sargsyan, K., Najm, H.N., Zádor, J.: Sella, an open-source automation-friendly molecular saddle point optimizer. Journal of Chemical Theory and Computation 18, 6974–6988 (2022) https://doi.org/10.1021/acs.jctc.2c00395
  69. [69]O’Boyle, N.M., Banck, M., James, C.A., Morley, C., Vandermeersch, T., Hutchison, G.R.: Open babel: An open chemical toolbox. Journal of Cheminformatics 3, 1–14 (2011) https://doi.org/10.1186/1758-2946-3-33
  70. [70]Schreiner, M.: Transition1x. figshare (2022). https://doi.org/10.6084/m9.figshare.19614657
  71. [71]Grambow, C.A., Pattanaik, L., Green, W.H.: Reactants, products, and transition states of elementary chemical reactions based on quantum chemistry. Zenodo (2020). https://doi.org/10.5281/zenodo.3581266
  72. [72]Simm, G.N., Reiher, M.: Context-driven exploration of complex chemical reaction networks. Journal of Chemical Theory and Computation 13, 6108–6119 (2017) https://doi.org/10.1021/acs.jctc.7b00945
  73. [73]Hait, D., Estrada Pabon, J.D., Stohr, M., Martinez, T.J.: Locating ab initio transition states via geodesic construction on machine-learned potential energy surfaces. Journal of Chemical Theory and Computation 21(22), 11632–11644 (2025) https://doi.org/10.1021/acs.jctc.5c01221
  74. [74]Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th Annual International Conference on Machine Learning, pp. 41–48 (2009)DOI
  75. [75]Soviany, P., Ionescu, R.T., Rota, P., Sebe, N.: Curriculum Learning: A Survey (2022). https://arxiv.org/abs/2101.10382
  76. [76]Chen, R.T.Q.: torchdiffeq (2018). https://github.com/rtqichen/torchdiffeq

Paper details

vlnc:W_kd9t2dsr
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents