betaFeedback

WorksPhysics and Astronomy

Foundation Models for Discovery and Exploration in Chemical Space

Abstract

Accurate prediction of atomistic, thermodynamic, and kinetic properties from molecular structures underpins materials innovation. Existing computational and experimental approaches lack the scalability required to navigate chemical space efficiently. Scientific foundation models trained on large unlabelled datasets offer a path towards navigating chemical space across application domains. Here, we develop MIST, a family of molecular foundation models with up to an order of magnitude more parameters and data than prior works. Trained using a novel tokenizer, Smirk, which comprehensively captures nuclear, electronic, and geometric information, MIST learns a diverse range of molecules. MIST models have been fine-tuned to predict more than 400 structure-property relationships and have been shown to match or exceed state-of-the-art performance across diverse benchmarks, from physiology to electrochemistry. We demonstrate the ability of these models to solve real-world problems across chemical space from multiobjective electrolyte solvent screening to stereochemical reasoning for organometallics and mixture property prediction. The clearest demonstration of a foundation model is its ability to solve problems that were neither explicit targets of training nor central to the intentions of its developers. We identify olfactory perception mapping as such a problem, and show that MIST accurately predicted scent profiles and learned a hierarchical representation of olfactory space consistent with hyperbolic geometry. We formulated hyperparameter aware Bayesian neural scaling laws which eliminate the need for hyperparameter sweeps at every scale, making training large compute-optimal models feasible on a limited compute budget. The methods and findings presented here represent a significant step towards accelerating materials discovery, design, and optimization using foundation models.

Chemical space is extraordinarily vast, with estimates placing the number of small molecules on the order of 106010^{60} [1]. Moreover, it is heterogeneous, encompassing everything from simple organic molecules to inorganic salts to complex organometallics and mixtures of them [6, 7]. Navigating this space to find molecules with desirable chemical, thermal, biological, or perceptual properties has traditionally relied on capital- and time-intensive wet lab experimentation and/or computationally expensive first-principles simulations [2]. The cost of these methods becomes intractable at scale, limiting current exploration to only a minuscule fraction of chemical space [6, 8, 9].

Over the past decade, molecular machine learning has accelerated material development cycles, improved efficiency, and reduced costs [2]. However, most state-of-the-art molecular property prediction models still rely on labelled training data, which is sparse, scarce, and expensive to generate [2, 10, 11]. Consequently, these models are plagued by poor generalization, hampering their utility in real-world applications where novel molecules of interest differ substantially from those contained in curated training datasets [10].

The recent success of transfer learning in Natural Language Processing (NLP) [12] and Computer Vision (CV) [13] has led to interest in the development of large-scale Scientific Foundation Models (SciFMs) [3, 4, 14, 15]. Rather than train on scarce labelled datasets, SciFMs leverage large unlabelled scientific corpora to uncover representations broadly applicable to downstream tasks [4, 16], thereby alleviating the need for ever-larger labelled datasets. Molecular chemistry is especially well-positioned to benefit from NLP Foundation Model (FM) breakthroughs, as large libraries of text-encoded molecules are readily available [17]. Despite this, existing chemical FMs remain specialized to particular subdomains, thus failing to provide a general foundation on which to build [18], and with limited ability to generalize across chemical space [3, 4, 14]. There is a need for a single model capable of learning from billions of diverse molecules, while resolving isotopic, stereochemical, structural, and electronic nuances. Such a model can then discover unifying scientific principles and accelerate chemical design across application domains, from energy to olfaction.

Here, we introduce Molecular Insight SMILES Transformers (MIST), a family of molecular FMs with expanded coverage of chemical space trained using Smirk [19], a novel tokenization scheme which captures a comprehensive representation of molecular structure including nuclear, electronic, and geometric features. Motivated by scaling trends in NLP [20, 21, 22], the largest MIST models were trained with an order of magnitude more parameters and data than prior work [3, 4, 14], matching or exceeding the state-of-the-art across diverse chemical benchmarks. We demonstrate MIST’s efficacy in solving problems across chemical space: multiobjective electrolyte solvent screening, organometallic stereochemical reasoning, isotope half-life prediction, mixture property prediction and olfactory perception mapping. Moreover, we show that generalizable chemical concepts are learned by the model during training. We examine every layer of the model for encoded chemical knowledge and find identifiable patterns, such as Hückel’s aromaticity rule and Lipinski’s Rule of Five (RO5), not explicitly labelled in the training data. These findings position MIST as a powerful tool for discovery and exploration in chemical space.

To train MIST compute-optimally, we augment the existing formulation of neural scaling laws [20] to derive new insights about the importance of data quality from the scaling exponents, which we expect to inform the data curation for FMs and SciFMs across scientific disciplines. Neural scaling laws introduce the notion of a compute-optimal frontier with respect to dataset size, model size, and compute budget [20]. However, current formulations assume all other hyperparameters (e.g., learning rate) have already been optimally tuned [20, 21]. Here, we introduce penalty terms to explicitly model the impact of off-optimal hyperparameter selection on data, model, and compute scaling. We parameterize these penalized neural scaling laws using Bayesian parameter estimation, providing a statistically rigorous framework for identifying scaling exponents. We show how the fitted scaling of data versus model, D∝NβˉD \propto N^{\bar{\beta}}, can provide a rich signal of the model’s operating regime and potential limitations of the pretraining corpus.

MIST establishes a paradigm for FM-guided, accelerated design of molecular materials and mixture formulations. Spanning chemical space from organics to organometallics and encompassing nuclear, electronic, and geometric diversity, the MIST models are poised to deepen our understanding of molecules and materials. Simultaneously, this work provides a path towards economizing and optimizing the training of future SciFMs by augmenting the existing compute-optimal neural scaling laws with hyperparameter penalties and linking data quality to scaling efficiency.

Results

Molecular FMs in the MIST family range in size from a few million up to 1.8 billion parameters, and are pretrained on the Simplified Molecular Input Line Entry System (SMILES) representations of up to 6B molecules. The two primary pretrained encoders (Methods table 1 and Supplementary section A) discussed in this paper are MIST-28M (28 million parameters trained on 245 million molecules) and MIST-1.8B (1.8 billion parameters trained on 2 billion molecules). MIST models use an encoder-only transformer [23, 24] architecture and were pretrained using the Masked Language Modeling (MLM) objective [12] (Methods section 3.1) on the Enamine REALSpace Dataset [17] (Supplementary section A), which consists of synthetically accessible organic molecules.

MIST models have been fine-tuned (Methods section 3.2) on over 400 molecular and formulation property prediction tasks (Supplementary section C), and demonstrate state-of-the-art performance across many chemical machine learning benchmarks (Extended Data Table S2, Table S3 and Table S5). Fine-tuned MIST models consist of the pretrained MIST encoder followed by a task network. The task networks (except for the specialized mixture task networks described in § 0.1.1.4) are two-layer Multi-Layer Perceptrons (MLPs).

log(global_throughput)generation_efficiency
(1)(2)
(Intercept)7.043***1.410**
(0.063)(0.494)
log(ranks)0.838***-0.012***
(0.007)(0.003)
Batch Size0.004***
(0.000)
(Batch Size)2^2-0.000***
(0.000)
Num. Fragments-0.000
(0.000)
Epoch Size0.053-1.503
(0.060)(0.993)
(Epoch Size)2^20.644
(0.496)
Limit Ref. Fragments-0.059***
(0.007)
Limit Inv. Fragments0.039***
(0.007)
NN707707
Degrees of Freedom701701
R2R^20.9650.206
MAE0.0960.043
RMSE0.1270.063

Table S2. Linear models used to guide performance tuning of our high-throughput screening workflow. Asterisks denote coefficient significance: * for p≤0.05p \le0.05, ** for p≤0.01p \le0.01, and *** for p≤0.001p \le0.001. Descriptions of each parameter are provided in section D.1

#HOMO (eV)LUMO (eV)Gap (eV)Melt (°C)Boil (°C)
1-9.82.212.1-9108
2-9.82.212.0-2110
3-10.01.811.9-1288
4-9.91.911.8-6282
5-9.62.111.8-13116
6-9.62.111.7-14118
7-10.01.811.7-4286
8-9.52.111.7-7120
9-9.71.711.4-6278
10-8.72.511.3-11159
11-9.12.111.2-8143
12-8.82.311.1-31174
13-8.82.311.1-58138
14-9.02.111.1-38156
15-9.02.111.1-20162
16-8.92.111.1-33166
17-9.51.511.1-4287
#HOMOLUMOGapMeltBoil
18-8.91.910.9-8198
19-9.41.510.9-2790
20-9.21.610.9-26127
21-9.31.510.8-6284
22-8.62.210.8-59106
23-8.22.410.7-12276
24-8.22.410.6-106113
25-9.31.210.6-3289
26-8.02.510.5-66165
27-8.02.310.4-58225
28-8.22.110.4-72115
29-8.91.410.4-44104
30-8.12.210.3-49167
31-7.82.410.2-32226
32-7.92.310.1-74163
33-8.02.010.1-94120
34-9.01.010.1-26160
35-7.42.39.8-1256
36-7.42.49.8-84134
37-7.52.19.7-73165
38-7.42.39.7-77142
39-7.81.79.6-10786
40-7.12.59.5-2272
41-7.22.39.5-116112
42-7.02.49.5-14242
43-7.12.29.4-6304
44-7.12.19.2-18229
45-8.80.49.2-24316
46-7.11.99.1-59244
47-7.12.09.1-84138
48-7.02.09.0-27302
49-8.80.18.9-35312
50-8.7-0.48.3-7099
51-7.30.88.1-84146
52-7.80.27.9-64179
53-7.30.47.8-36266
54-7.30.57.8-39229
55-7.30.57.8-62185
56-7.10.67.7-7336
57-7.20.57.7-43251
58-7.9-0.27.7-70151
59-7.10.57.7-39332
60-7.10.67.7-78151
61-7.10.57.6-54272
62-7.20.37.6-46231
63-7.00.57.5-66185

Table S3. Pareto Front Candidate Electrolytes and computed properties using fine-tuned MIST-28M models. Molecular structures (Indexed by #) are shown in figs. S13 to S17.

nnLnL_nH^γ\hat{H}_\gammaHH
Generated1,9367,600.45707.857.8
Electrolytes105681.72962.431.4

Table S5. Estimated Rényi entropy H^γ\hat{H}_\gamma (eq. S4) and average molecular surprise HH, akin to Shannon entropy, for generated molecules and reference electrolytes. We observe a reversal in the ordering of HH and H^γ\hat{H}_\gamma because the two metrics measure different phenomena: H^γ\hat{H}_\gamma captures the entropy of the generation process, while HH evaluates the entropy of the generated molecules.

To train MIST models on SMILES sequences for a diverse range of molecules, the Smirk [19] tokenization algorithm was developed. In contrast to existing tokenizers for chemistry, Smirk captures nuclear, electronic, and geometric features enabling SciFMs to learn a rich representation of molecular structure [19]. This capacity enables fine-tuning on chemistries including radioactive isotopes and organometallic complexes. Interestingly, we find that despite the limited diversity of the pretraining dataset (discussed further in § 0.1.3 and Supplementary section P), the downstream performance of MIST models benefits from the pretraining stage, even on tasks involving far more diverse chemical species (§ 0.1.1.3 and Supplementary section A.5).

We developed and leveraged hyperparameter-penalized neural scaling laws (§ 0.1.3) to remain on the compute-optimal frontier while pretraining MIST models. We reduced the cost of fitting scaling laws, using Bayesian parameter estimation and design of experiments [25]. This reduced our compute cost by an estimated 10 petaflop-days, an order of magnitude reduction, relative to the grid search widely used to fit neural scaling laws [21, 26, 27].

MIST has Applications Across Chemical Space

The core advantage of a foundation model is that it can be adapted to a wide range of downstream tasks given a small number of labelled examples [18]. We demonstrate the MIST models’ efficacy as SciFMs by fine-tuning variants of MIST to predict over 400 properties — including quantum mechanical, thermodynamic, biochemical, and psychophysical properties — from a molecule’s SMILES representation (Figure 1a and Supplementary section C.1). The MIST encoders are fine-tuned on labelled datasets as small as ∼180\sim180 examples. The encoders are fine-tuned on single molecule property prediction (classification and regression) tasks by attaching a small two-layer MLP (Methods section 3.2). They are fine-tuned on mixture property prediction tasks using a physics-informed task network architecture (Methods section 3.4). As we show in the following sections, these fine-tuned models unlock problem-solving capabilities across vast regions of chemical space at multiple scales, from single-molecule electrolyte solvents to large chiral organometallic structures and complex multi-component mixtures. Notably, in olfaction—a stringent test of transfer beyond the intended scope of pretraining—we experimentally validated MIST’s ability to resolve structure–odour discordances, demonstrating that its learned representations capture perceptual relationships not evident from structure alone.

Figure 1. MIST models molecules across chemical space, accelerating a broad range of materials design tasks. (a) Molecular Insight SMILES Transformers (MIST) is a family of molecular FMs which match or exceed state-of-the-art performance across diverse benchmarks, from physiology to quantum chemistry. The models solve real-world problems across chemical space — ranging from multiobjective electrolyte solvent screening and olfactory perception mapping to stereochemical reasoning for organometallic compounds and mixture property prediction. (b) The MIST models are encoder-only transformers [23, 24] pretrained using Masked Language Modeling and a novel tokenization scheme, Smirk [19], to learn informative vector embeddings of molecules from their SMILES representations. During fine-tuning, the molecular embeddings are processed by small task networks, to predict properties of interest.

High-Throughput Screening for Electrolytes

Energy storage and conversion devices require electrolyte materials that enable functionality for various applications such as electric mobility, grid storage and chemicals production [30, 31]. Finding new molecules that meet the full spectrum of physicochemical property criteria for use in energy storage devices — electrochemical stability, a wide operating liquid range, and good transport properties — is a challenging multi-objective optimization problem [31]. High-throughput computational screening has emerged as an efficient and cost-effective strategy to accelerate electrolyte discovery, reducing the number of wet-lab experiments required [32]. Fine-tuned MIST models provide fast and accurate predictions for a broad range of electrolyte properties, making them particularly effective in these pipelines.

We built a high-throughput screening pipeline for electrolyte design (Methods section 3.3) using fine-tuned MIST-28M models. The screening pipeline begins with a seed inventory from which novel molecules are generated using FASMIFRA [33]. The generated molecules are screened against the specified criteria

using MIST models fine-tuned to predict the relevant properties. We demonstrate the efficacy of this pipeline in two settings: first, a broad search for solvents with improved electrochemical and thermophysical stability (Supplementary section D.2) and second, a more tightly constrained problem of solvent discovery for lithium–air batteries (Supplementary section D.3).

Electrochemical stability was evaluated using a MIST-28M model fine-tuned on the QM9 dataset [34] and thermal stability with models fine-tuned on a Characteristic Temperature (CT) dataset to predict melting and boiling points. For lithium–air screening, additional models were fine-tuned to predict pKapK_a in Dimethyl Sulfoxide (DMSO) (Supplementary section C.1.7), Kamlet-Taft (KT) solvent parameters (Supplementary section C.1.3) and normalized Dimroth–Reichardt solvent polarity (ETNE_T^N, section C.1.6). Across a curated set of 92 electrolyte solvents, these models recovered expected property relationships, including the established linear correlation between flash point and boiling point provided by Hess et al. [35] (section D.2). Selected predictions on generated molecules were further checked using Density Functional Theory (DFT) as discussed in Supplementary section D.2.2.

Applied to the joint electrochemical and thermophysical screen, this workflow evaluated 90 million molecules in 8 hours on 8 NVIDIA A100 GPUs and identified 63 Pareto-optimal solvent candidates. Critically, the generated candidates improved upon both the electronic and thermal properties of commonly used electrolytes (section D.2).

Analyzing molecules on the Pareto front, we found that generating and evaluating novel molecules during a screening campaign is important. We evaluated the novelty of a molecule using eq. (1), which we call the molecular surprise:

−∑i=0Nln⁡exp⁡Q(xi)∑x∈Vexp⁡Q(x)(1)-\sum_{i=0}^{N} \ln\frac{\exp Q(x_i)}{\sum_{x \in V} \exp Q(x)} \tag*{(1)}

where Q(xi)Q(x_i) gives the model’s logit score for token ii and the denominator marginalizes over all other tokens in the vocabulary VV. Access to more surprising molecules yielded several Pareto front candidates with improved electrochemical stability (Figure S24), highlighting the importance of generating and evaluating novel molecules during a screening campaign.

Figure S24. “Surprising” generated molecules yielded improved performance characteristics, for example, a wider electrochemical stability window (HOMO-LUMO Gap).

Pareto front molecules (Supplementary figs. S13 to S17) were consistent with established electrolyte heuristics — 57% were fluorinated and esters were the most common functional group [31]. Additionally, several Pareto-optimal candidates also exhibit mixed halogenation, including fluorine/bromine substitution in 1,8-dibromoperfluorooctane (Molecule 4, Supplementary fig. S13) and fluorine/chlorine substitution in (3S,5R)-pentachloro-nonafluorohexane (Molecule 29, Supplementary fig. S14). Their emergence suggests that the workflow recovers not only canonical fluorination-based design rules [36], but also an underexplored and chemically interpretable space in which partial replacement of fluorine with heavier halogens tunes the trade-off between electrochemical robustness and volatility. Follow-up analysis (Supplementary section D.2.3) showed that fluorine-dominant mixed halogenation can preserve electrochemical stability, whereas heavier halogens tend to raise boiling and flash points relative to fully fluorinated analogues.

We next applied the framework to the more challenging problem of solvent discovery for lithium–air batteries, where viable candidates must satisfy a narrow and mechanistically constrained design window. Imposing additional bounds associated with solution-mediated peroxide growth and resistance to hydrogen abstraction (Supplementary section D.3) yielded 390 Pareto-front candidate solvents. These molecules were enriched in chemically plausible Li−O2\mathrm{Li}-\mathrm{O}_2 electrolyte motifs, including haloalkanes, ethers, arenes, amines and fluorinated scaffolds. This indicates that the model identifies structurally interpretable solvent families rather than arbitrary admissible structures (Supplementary section D.3.3). This result provides a stringent demonstration that fine-tuned MIST models can translate literature-derived electrolyte design rules into a scalable molecular discovery workflow in a chemically meaningful regime.

Predicting Isotope Half-Life

Predicting the half-life of a radioactive isotope is a longstanding challenge in nuclear science. Accurate half-life predictions are crucial for understanding nuclear stability, the synthesis of heavy elements, and practical applications like nuclear medicine. Theoretical calculations of the half-life remain difficult because the decay constants depend nonlinearly on small errors in the binding energy and nuclear transition matrix elements. Competing decay channels (α\alpha, β±\beta^{\pm}, fission) further compound these uncertainties. Hence, the field has recently turned to machine learning based approaches [37].

The Smirk tokenizer’s ability to represent all isotopic variants for all elements enables us to fine-tune a MIST model for half-life prediction [19]. Projecting embeddings from the fine-tuned model using UMAP reveals structure in the embedding space (Figure 2b) explained by nuclide stability and dominant decay channels (Supplementary section G). We use this as a demonstration of the Smirk tokenizer’s ability to parse isotopic information in SMILES (for example [2H] and [13C]). Although a SMILES-based model is not necessary for this particular task, this shows MIST’s ability to tackle molecular property prediction problems with isotopic variation.

Figure 2. MIST rapidly explores chemical space. (a) MIST-28M accurately predicts quantum, chemical, and thermodynamic descriptors for electrolyte design — including orbital energy levels and solvatochromic parameters — when fine-tuned on labelled datasets as small as ∼180\sim180 samples. Using fine-tuned MIST-28M models, we identified 390 solvent molecules for lithium-air batteries on the Pareto front, improving stability against hydrogen abstraction and encouraging solution-mediated peroxide growth relative to our reference set. (b) A Uniform Manifold Approximation and Projection (UMAP) projection of embedding vectors from a MIST-28M variant fine-tuned on the half-lives of radioisotopes reveals nuclide clusters organized by dominant decay channels. (c) Parity plots for MIST-28M models, fine-tuned on transition metal organometallic complexes, with annotated examples of stereoisomer pairs. Stereochemical nuances are accurately captured by the Smirk tokenizer, enabling MIST to differentiate between isomers. (d) To fine-tune MIST models on binary mixture properties, we use a specialized permutation-invariant task network. This task network uses MIST embeddings e⃗i\vec{e}_i to compute the linear mixing PLP_L and excess PEP_E contributions to the mixture property PmixP_{\mathrm{mix}}. (e) Electrolyte ionic conductivity is predicted using a physics-informed task network from a concentration-weighted sum of MIST embeddings. (f) MIST mixture property models accurately predict excess properties, including for challenging mixtures not captured by similarity-based descriptors [28]. (g) Identifying additives with high relative excess at small concentrations is critical during electrolyte formulation [29]. MIST models accurately estimate the composition at the excess extrema. (h) MIST models learn physically consistent trends in ionic conductivity with salt mole fraction (xLix_{\mathrm{Li}}), correctly capturing the expected decay outside the training data range (xLi>0.15x_{\mathrm{Li}} > 0.15, highlighted in yellow).

Modelling Organometallic Complexes

Transition metal complexes are fundamental for a wide range of applications, from biochemistry [38] to photochemistry [39]. Understanding the electronic structure and properties of transition metal organometallics is key to their functionality [39]. After fine-tuning MIST on a dataset of transition metal complexes [40], the model accurately predicts quantum mechanical properties, with MIST-28M and MIST-1.8B achieving validation R2R^2 scores of 0.76 and 0.80, respectively. Notably, the Smirk tokenizer [19] tokenizes the relevant chiral tags (e.g., @SP1 versus @SP3), enabling MIST to differentiate between non-tetrahedral isomers (Figure 2c). By contrast, prior molecular FMs used tokenizers that obscured stereochemical features [19]. We also show that the MIST pretraining stage, while limited to organic molecules, improves performance on organometallic tasks (Supplementary section A.5).

Mixture Property Prediction

Mapping molecular properties with macroscale mixture properties, such as density, viscosity, and ionic conductivity, is challenging due to the combinatorial nature of the problem, complex non-ideal interactions between mixture components, and the scarcity of data [11]. Physics-based models, such as molecular dynamics simulations, can be used to study component interactions inside complex mixture systems. However, their computational cost can be prohibitive. Empirical relationships, such as the Arrhenius equation or Quantitative Structure-Activity Relationship (QSAR), which simplify intermolecular interactions, are widely used to inform formulation design. However, the accuracy of these models typically depends on adequate parameterization with experimental data, and their ability to generalize across chemical space is limited [11].

Machine learning based approaches have emerged as a feasible and scalable solution for mixture property prediction. Building on prior work [11, 41], we extend MIST’s molecular property prediction capabilities to mixture property prediction using downstream task networks informed by chemical thermodynamics. Fine-tuned MIST models achieve state-of-the-art performance on existing benchmark datasets for excess molar volume, excess molar enthalpy, and ionic conductivity (Extended Data Table S5), outperforming purpose built Graph Neural Networks (GNNs) including those which require 3D geometrical information [11].

Excess Property Prediction. Excess properties measure the deviation of mixture properties from ideal mixing behaviour. These deviations arise from molecular interactions between components [42]. Understanding these deviations is particularly relevant when identifying additives that can meaningfully impact performance metrics at low concentrations [29]. Mixture properties can be decomposed into a linear mixing and an excess term. For a binary mixture of substances Si∣i∈{1,2}S_i \mid i \in\{1,2\}, a mixture property PmixP_{\mathrm{mix}} can be expressed as:

Pmix=∑i=12xiPi⏟PL (linear mixing)+∑j=1nΘj⋅fj(x1)⏟PE (excess term)(2)P_{\mathrm{mix}} = \underbrace{\sum_{i=1}^{2} x_i P_i}_{P_{\mathrm{L}}\ \text{(linear mixing)}} + \underbrace{\sum_{j=1}^{n} \Theta_j \cdot f_j(x_1)}_{P_{\mathrm{E}}\ \text{(excess term)}} \tag*{(2)}

where x1x_1 and x2x_2 are the mole fractions of the two components in the mixture, PiP_i is the property of the pure substance SiS_i, fjf_j is the jthj^{\mathrm{th}} basis of a polynomial and Θj\Theta_j is the corresponding empirically fit coefficient. The excess term is commonly modelled using a Redlich–Kister polynomial basis to enforce PE=0P_{\mathrm{E}} = 0 when x1=0x_1 = 0 or x2=0x_2 = 0 [11, 43].

Our binary mixture property prediction task network is informed by the thermodynamic statement of excess properties (eq. (2) and Figure 2d), and preserves key physical principles: PE=0P_E = 0 in the degenerate case of a pure compound, and excess properties are permutation invariant. We fine-tune MIST to predict PiP_i and Θj\Theta_j of eq. (2) using a permutation-invariant task network (Figure 2d) that produces smooth, physically consistent excess-property curves (Methods section 3.4).

We use the MIST model to accurately estimate the “excess asymmetry” (Figure 2g) in density and molar volume. We define the “excess asymmetry” as the deviation of the composition of the excess extremum from that of a regular solution model, that is 0.5. This prediction enables identification of additives with high relative excess at low concentrations, which is critical for additive design in electrolytes and liquid formulation problems [29]. The MIST model accurately identifies the emergence of excess properties in mixtures that are not well modelled by existing similarity-based descriptors. We demonstrate this by identifying a set of mixtures that disagree with the negative correlation between structural similarity and excess magnitude observed in prior work [28] (Supplementary fig. S71a) and show that the model accurately predicts the excess properties of these mixtures (Figure 2f).

Mapping Olfactory Perception

Multi-Component Electrolyte Property Prediction. MIST models were fine-tuned to predict the ionic conductivity of electrolytes (salts in ternary solvent systems) [11]. A specialized multi-component property prediction task head is used to enable this. The task head constructs a mixture embedding, e⃗mix\vec{e}_{mix}, as the sum of the component embeddings e⃗i\vec{e}_i weighted by their mole ratios in the mixture xix_i:

e⃗mix=∑i=1nxie⃗i(3)\vec{e}_{mix} = \sum_{i=1}^{n} x_i \vec{e}_i \tag*{(3)}

This mixture embedding is used to predict two sets of coefficients. The first set of coefficients parameterize the Vogel-Fulcher-Tammann (VFT) relation [44], which the model uses to learn the dependence of ionic conductivity on temperature. We show that the VFT relation accurately describes the temperature dependence seen in electrolytes in Supplementary section I.1. The second set of coefficients parameterize an empirical correction term used by the model to learn the dependence of ionic conductivity on concentration (Methods section 3.4.2).

These physics-informed relationships allow the model to generalize to higher salt mole fractions and concentrations than those seen during training; as shown in Figure 2h, the model correctly captures the expected decay outside the training data range (xLi>0.15x_{Li} > 0.15). Using the learnt VFT pseudo-activation energy EaE_a and Vogel temperature T0T_0, we compute the associated excess quantities (deviation from linear mixing across solvent fractions, see Supplementary sections I.3 and I.4). Sweeping along the salt and solvent concentration axes, we disentangle solvent–solvent and salt–solvent effects on ion transport across the ternary composition space. The model recovers the expected trend that linear carbonate-rich solvents exhibit smaller EaE_a and T0T_0 (Supplementary figs. S73 to S76 and S89 to S92), whereas viscous cyclic carbonates (Ethylene Carbonate (EC), Propylene Carbonate (PC), and Fluoroethylene Carbonate (FEC)) show the opposite behaviour (Supplementary figs. S79, S80, S95 and S96). The ternary excess surfaces of EaE_a and T0T_0 reveal nonideal mixing, with a positive excess near linear carbonate rich corners (Supplementary figs. S81, S82, S97 and S98) and a negative excess in moderate cyclic carbonate compositions, where viscosity decreases without a corresponding increase in association. The model captures a subtle yet important design principle: adding small amounts of linear carbonates to cyclic-carbonate mixtures reduces the viscosity while maintaining high ionic dissociation. The MIST model’s predictions suggest an interesting effect of the salt in the electrolyte, LiPF6_6 exhibits higher pseudo-activation steepness (Ea/TE_a/T) than LiTFSI, indicating stronger temperature-activated barriers for LiPF6_6 (Supplementary section I.5). Together, the learnt VFT coefficients and their excesses provide mechanistic composition-resolved guidance for selecting solvent ratios and salts that reduce fragility and improve cold-temperature performance. Supplementary section I provides additional details and results for this analysis.

The clearest demonstration of a foundation model is its ability to solve problems that were neither explicit targets of training nor central to the intentions of its developers [18]. Predicting the odour of a molecule

remains a longstanding challenge in chemistry, psychophysics and neuroscience because the mapping from molecular structure to perception is highly nonlinear, discontinuous and variable across individuals [45, 46, 47]. Minimal chemical modifications can produce large perceptual changes, whereas structurally dissimilar molecules can evoke near-identical odours [5]. We therefore identified olfaction as a particularly difficult test of MIST: it lies outside the applications motivating model development, yet success in this domain would suggest that MIST has learned molecular representations that transfer beyond the tasks targeted during pretraining [48, 49].

After fine-tuning MIST-28M on a dataset of ∼\sim 5,000 molecules curated by Lee et al. [5], we find the MIST model exceeds the accuracy of the median rater on an expert panel and existing state-of-the-art models [5]. The fine-tuned model has an Area Under the Receiver Operating Characteristic (AUROC) of 0.915, improving on the 0.894 AUROC reported by Lee et al. [5] for their GNN.

The MIST model’s performance scales with the chemical specificity of scent descriptors. Scents associated with particular functional groups are highly separable; for example, fruity scents caused by the presence of an ester group (“banana”, “apple” and “cognac”) all have AUROCs above 0.9. The model accurately captures dramatic changes in odour caused by small changes in molecular structure; for example, it accurately labels acetone CC(=O)C\mathrm{CC(=O)C} as “ethereal” and that a sulfur substitution (thioacetone CC(=S)C\mathrm{CC(=S)C}) leads to a “sulfurous” and “meaty” odour (Supplementary fig. S65d). We validate the MIST model’s robustness to discontinuities in structure-odour mapping by using it to discover previously uncharacterized “discordant” triplets (Figure 3a, Supplementary section F.1). Discordant triplets are triplets of molecular odorants in which the structurally similar pair is not the perceptually similar pair [5]. We evaluated structural similarity using Tanimoto similarity and perceptual similarity using the Euclidean distance between predicted MIST odour profiles. The experimentally determined labels for these triplets were not publicly available at the time of prediction; we subsequently obtained them through direct correspondence with Lee et al. providing a prospective, blinded evaluation. The MIST model correctly resolved 64.3% of the discordant triplets (Supplementary fig. S66e), outperforming the GNN of Lee et al. [5], which resolved 50% (Supplementary fig. S66c).

MIST structure–odour mapping, hierarchical scent relationships, hyperbolic olfactory space, and multi-component odour mixturesVector (PDF)
e⃗1,mix=∑iw1,ie⃗1,i\vec{e}_{1,\mathrm{mix}}=\sum_i w_{1,i}\vec{e}_{1,i}
e⃗2,mix=∑kw1,ke⃗1,k\vec{e}_{2,\mathrm{mix}}=\sum_k w_{1,k}\vec{e}_{1,k}
∥e⃗1,mix−e⃗2,mix∥M\left\lVert\vec{e}_{1,\mathrm{mix}}-\vec{e}_{2,\mathrm{mix}}\right\rVert_M

Figure 3. MIST captures discontinuities in structure–odour mapping and the hierarchical geometry of olfactory space. (a) Structure alone often fails to explain odour similarity: across 41 odour-discordant triplets, the experimentally closer-smelling molecule is frequently not the structurally closest one. Using Euclidean distance between perceptual profiles predicted by MIST, we are able to correctly identify 64.3% of these discordant triplets, outperforming the prior state of the art (50.0%). (b) Hierarchical clustering of MIST scent logits recovers human-interpretable scent relationships, including grouped savoury and fruity descriptors and a distinct “odourless” cluster. (c) Shell-constrained Hyperbolic Distance Recovery and Approximation (HYDRA) embedding visualizes the learned odour representation in hyperbolic space. Circles/rectangles fall on the near/far side of the sphere. The RGB color scales were proportional to the XYZ coordinates of points. (d) The MIST model was extended to predict perceptual distances between multi-component odour mixtures.

We observe that the model learns a hierarchical clustering of scents (Figure 3b, Methods section 3.7) consistent with human perceptual organization; for example, “beefy”, “meaty”, “roasted”, “savory” all form a cluster, fruity scents (such as “strawberry”, “berry”, “plum” and “grape”) form another and notably “odourless” molecules are clustered alone. The existence of this hierarchical structure is consistent with the model having learned the hyperbolic geometry of the olfactory space [50]. To validate this, we modeled the olfactory profiles in hyperbolic and Euclidean space. We compared the empirical Betti curves of the MIST logit similarity structure with those generated from best-fit Euclidean and hyperbolic model spaces using ALBATROSS [51, 52] (Methods section 3.7). The MIST logit geometry was well fit by a hyperbolic shell model (p=1.0p = 1.0, χ2=0\chi^2 = 0) (Supplementary section F.3). Together, these results indicate that the learned olfactory representation is statistically consistent with a low-dimensional hyperbolic geometry. This finding is consistent with prior work showing that both natural odour co-occurrence statistics and human olfactory perceptual spaces are well described by low-dimensional hyperbolic geometry [50]. We visualized the learned perceptual representation as a hyperbolic shell (Figure 3c) using shell-constrained HYDRA embeddings [51, 53] (Methods section 3.7).

Finally, we extend the model’s capabilities beyond predicting the odour of a single molecule to comparing the odour profiles of multi-component mixtures. We fine-tuned MIST to predict experimentally measured perceptual distances ranging from 0 (indistinguishable odour) to 1 (maximally distinct odours) for pairs of mixtures with variable numbers of components (see Supplementary section C.1.12 for dataset details). We use a weighted sum of single molecule embeddings (with learned weights) to obtain a mixture embedding (Figure 3d). The fine-tuned MIST achieved a Pearson correlation of 0.62±0.050.62 \pm0.05 on a held-out validation split, demonstrating comparable performance to an ensemble of top-performing models submitted to the Dialogue for Reverse Engineering Assessment and Methods (DREAM) Olfaction Challenge 2024, which reported a Pearson correlation of 0.57 [54].

Learning a Chemically Meaningful Representation

The potential of SciFMs, such as MIST, for scientific discovery extends beyond their raw predictive accuracy. For exploration, scientists need to be able to extract interpretable features aligned with scientific concepts and relationships which can be understood and steered to generate hypotheses or diagnose failure modes [16,

[55]. In NLP and CV, mechanistic interpretability has revealed feature vectors within FMs correlated with semantically meaningful concepts across modalities and languages [56, 57]. Yet comparable efforts for SciFMs remain limited [16]. Prior SciFM work has examined attention maps [3, 14] and low-dimensional projections of embedding space [3, 4, 58] to probe internal behaviour.

We systematically search for interpretable, concept-aligned features across all layers, and we demonstrate that MIST models consistently encode chemical concepts that generalize beyond their training objectives (Figure 4). Such concept-aligned features can be used to enable the discovery of new material design rules and robust generalization to novel tasks, thereby positioning MIST as a practical tool for hypothesis generation and systematic discovery across chemical space [16].

Figure 4. Interpretability analysis reveals MIST’s potential as a robust tool for exploration and discovery. (a) Interpretable chemical features were extracted from every stage of MIST models — token embeddings to downstream predictions. (b) A t-SNE projection of token embeddings shows dataset-specific structure. All three datasets shown cluster the halogens and digits. The transition metal Quantum Mechanics dataset (tmQM) forms clusters for the left and right transition metal block. FreeSolv orders the frequently seen p-block elements (Si P N O S I Br F and C) near the bottom of the figure. (d) & (c) UMAP projections reveal that the pretrained embeddings encode ring aromaticity and Polycyclic Aromatic Hydrocarbons classes, despite the model never training on labelled examples of these classes, suggesting the existence of chemically coherent structures in MIST’s embedding space. (e) Using fine-tuned MIST-28M variants, we recover expected trends (increasing and converging) in HOMO energies and phase transition temperatures (Melt & Boil) with carbon chain length.

Interpreting Token Embeddings

References

  1. [1]Christopher M. Dobson. “Chemical Space and Biology”. In: *Nature* 432.7019 (Dec. 1, 2004), pp. 824–828.DOI
  2. [2]Aashutosh Mistry et al. “How Machine Learning Will Revolutionize Electrochemical Sciences”. In: *ACS Energy Lett.* 6 (Mar. 23, 2021), pp. 1422–1431.DOI
  3. [3]Ross, J., & al., . et . (2022). Large-Scale Chemical Language Representations Capture Molecular Structure and Properties. Nat Mach Intell, 4.12, 1256–1264.
  4. [4]Soares, E., & al., . et . (2024). A Large Encoder-Decoder Family of Foundation Models For Chemical Language. 2407.20267
  5. [5]Lee, B. K., & al., . et . (2023). A Principal Odor Map Unifies Diverse Tasks in Olfactory Perception. Science, 381.6661, 999–1006.
  6. [6]Kirkpatrick, P., & Ellis, C. (2004). Chemical Space. Nature, 432.7019, 823–823.
  7. [7]Jiyeon Kim et al. “Up–down Approach for Expanding the Chemical Space of Metal–Organic Frameworks”. In: *Nature Synthesis* 3.12 (Dec. 2024), pp. 1518–1528.DOI
  8. [8]Lipinski, C., & Hopkins, A. (2004). Navigating Chemical Space for Biology and Medicine. Nature, 432.7019, 855–861.
  9. [9]Gao, W., Luo, S., & Coley, C. W. (2025). Generative AI for Navigating Synthesizable Chemical Space. Proceedings of the National Academy of Sciences, 122.41, e2415665122.
  10. [10]Naohiro Fujinuma et al. “Why Big Data and Compute Are Not Necessarily the Path to Big Materials Science”. In: *Commun Mater* 3.1 (Aug. 30, 2022), p. 59.DOI
  11. [11]Shang Zhu et al. “Differentiable Modeling and Optimization of Non-Aqueous Li-based Battery Electrolyte Solutions Using Geometric Deep Learning”. In: *Nat Commun* 15.1 (Oct. 5, 2024), p. 8649.DOI
  12. [12]Devlin, J., & al., . et . (2019). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. 1810.04805
  13. [13]Radford, A., & al., . et . (2021). Learning Transferable Visual Models From Natural Language Supervision. 2103.00020
  14. [14]Seyone Chithrananda, Gabriel Grand, and Bharath Ramsundar. *ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction*. Oct. 23, 2020. arXiv: 2010.09885 [physics, q-bio]. url: http://arxiv.org/abs/2010.09885. Pre-published.
  15. [15]Xiang, J., & al., . et . (2025). A Vision–Language Foundation Model for Precision Oncology. Nature, 638.8051, 769–778.
  16. [16]O’Neill, C., & al., . et . (2024). Towards Interpretable Scientific Foundation Models: Sparse Autoencoders for Disentangling Dense Embeddings of Scientific Concepts. Neurips 2024 Workshop Foundation Models for Science: Progress, Opportunities, and Challenges.
  17. [17]Ltd., E. (2024). REAL Space.
  18. [18]Bommasani, R., & al., . et . (2022). On the Opportunities and Risks of Foundation Models. 2108.07258
  19. [19]Alexius Wadell, Anoushka Bhutani, and Venkatasubramanian Viswanathan. “Tokenization for Molecular Foundation Models”. In: *Journal of Chemical Information and Modeling* (Jan. 23, 2026).DOI
  20. [20]Hoffmann, J., & al., . et . (2022). Training Compute-Optimal Large Language Models. 2203.15556
  21. [21]Kaplan, J., & al., . et . (2020). Scaling Laws for Neural Language Models. 2001.08361
  22. [22]Bi, X., & al., . et . (2024). DeepSeek LLM: Scaling Open-Source Language Models with Longtermism. 2401.02954
  23. [23]Vaswani, A., & al., . et . (2017). Attention Is All You Need. 1706.03762
  24. [24]Liu, Y., & al., . et . (2019). RoBERTa: A Robustly Optimized BERT Pretraining Approach.
  25. [25]Izabella Surowiec et al. “Generalized Subset Designs in Analytical Chemistry”. In: *Anal. Chem.* 89.12 (June 20, 2017), pp. 6491–6497.DOI
  26. [26]Margaret Li, Sneha Kudugunta, and Luke Zettlemoyer. *(Mis)Fitting: A Survey of Scaling Laws.* 2025. arXiv: 2502.18969 [cs.LG].arXiv
  27. [27]Hägele, A., & al., . et . (2024). Scaling Laws and Compute-Optimal Training Beyond Fixed Training Durations. 2405.18392
  28. [28]Celia Kelly et al. “Excess Density as a Descriptor for Electrolyte Solvent Design”. In: *The Journal of Chemical Physics* 162.6 (Feb. 10, 2025), p. 064301.DOI
  29. [29]Austin D. Sendek and Venkatasubramanian Viswanathan. “Artificial Intelligence for Electrolyte Design: Going Beyond the Molecular Paradigm”. In: *Electrochem. Soc. Interface* 34.2 (June 1, 2025), p. 39.DOI
  30. [30]Xu, K. (2004). Nonaqueous Liquid Electrolytes for Lithium-Based Rechargeable Batteries. Chem. Rev., 104(10), 4303–4418.
  31. [31]Kang Xu. “Electrolytes and Interphases in Li-Ion Batteries and Beyond”. In: *Chem. Rev.* 114.23 (Dec. 10, 2014), pp. 11503–11618.DOI
  32. [32]Lei Cheng et al. “Accelerating Electrolyte Discovery for Energy Storage with High-Throughput Screening”. In: *Journal of Physical Chemistry Letters* 6.2 (2015), pp. 283–291.DOI
  33. [33]Francois Berenger and Koji Tsuda. “Molecular Generation by Fast Assembly of (Deep)SMILES Fragments”. In: *J Cheminform* 13.1 (Dec. 2021), p. 88.DOI
  34. [34]Raghunathan Ramakrishnan et al. “Quantum Chemistry Structures and Properties of 134 Kilo Molecules”. In: *Sci Data* 1.1 (Aug. 5, 2014), p. 140022.DOI
  35. [35]Steffen Hess, Margret Wohlfahrt-Mehrens, and Mario Wachtler. “Flammability of Li-Ion Battery Electrolytes: Flash Point and Self-Extinguishing Time Measurements”. In: *J. Electrochem. Soc.* 162.2 (2015), A3084–A3097.DOI
  36. [36]Yumin Zhang and Venkatasubramanian Viswanathan. “Not All Fluorination Is the Same: Unique Effects of Fluorine Functionalization of Ethylene Carbonate for Tuning Solid-Electrolyte Interphase in Li Metal Batteries”. In: *Langmuir* 36.39 (Oct. 2020), pp. 11450–11466.DOI
  37. [37]Amir Jalili et al. “$\alpha$-Decay Half-Life Predictions with Support Vector Machine”. In: *Scientific Reports* 14.1 (Dec. 28, 2024), p. 30776.DOI
  38. [38]Fabio Juliá. “Catalysis in the Excited State: Bringing Innate Transition Metal Photochemistry into Play”. In: *ACS Catal.* 15.6 (Mar. 21, 2025), pp. 4665–4680.DOI
  39. [39]Giacomo Morselli, Christian Reber, and Oliver S. Wenger. “Molecular Design Principles for Photoactive Transition Metal Complexes: A Guide for “Photo-Motivated” Chemists”. In: *J. Am. Chem. Soc.* 147.14 (Apr. 9, 2025), pp. 11608–11624.DOI
  40. [40]David Balcells and Bastian Bjerkem Skjelstad. “tmQM Dataset—Quantum Geometries and Properties of 86k Transition Metal Complexes”. In: *J. Chem. Inf. Model.* 60.12 (Dec. 28, 2020), pp. 6135–6146.DOI
  41. [41]Zohair, M., & al., . et . (2025). Chemical Foundation Model Guided Design of High Ionic Conductivity Electrolyte Formulations. http://arxiv.org/abs/2503.14878
  42. [42]Leon de Villiers Engelbrecht et al. “MD simulations explain the excess molar enthalpies in pseudo-binary mixtures of a choline chloride-based deep eutectic solvent with water or methanol”. In: *Frontiers in Chemistry* 10 (2022), p. 983281.DOI
  43. [43]Otto Redlich and A. T. Kister. “Algebraic Representation of Thermodynamic Properties and the Classification of Solutions”. In: *Ind. Eng. Chem.* 40.2 (Feb. 1948), pp. 345–348.DOI
  44. [44]K.L. Ngai. “Introduction to the Problems of Relaxation and Diffusion in Complex Systems”. In: Relaxation and Diffusion in Complex Systems. Ed. by K.L. Ngai. New York, NY: Springer, 2011, pp. 1–47.DOI
  45. [45]Andreas Keller et al. “Predicting Human Olfactory Perception from Chemical Features of Odor Molecules”. In: Science 355.6327 (Feb. 24, 2017), pp. 820–826.DOI
  46. [46]Malnic, B., & al., . et . (1999). Combinatorial Receptor Codes for Odors. Cell, 96.5, 713–723.
  47. [47]Rayane Achebouche et al. “Application of Artificial Intelligence to Decode the Relationships between Smell, Olfactory Receptors and Small Molecules”. In: Scientific Reports 12.1 (Nov. 5, 2022), p. 18817.DOI
  48. [48]U.S. Department of Energy, O. of S. (2023). 2024 INCITE Fact Sheet.Pdf.
  49. [49]Jolly, A. (2025). Building AI Foundation Models to Accelerate the Discovery of New Battery Materials.
  50. [50]Zhou, Y., Smith, B. H., & Sharpee, T. O. (2018). Hyperbolic Geometry of the Olfactory Space. Science Advances, 4.8, eaaq1458.
  51. [51]Laboratory, E. N. (2019). albatross: Fast filtration based geometry via stochastic sub-sampling.
  52. [52]Stier, A. J., & al., . et . (2025). ALBATROSS: Cheap Filtration Based Geometry via Stochastic Sub-Sampling. 2509.03681
  53. [53]Martin Keller-Ressel and Stephanie Nargang. Hydra: A method for strain-minimizing hyperbolic embedding of network- and distance-based data. 2019. arXiv: 1903.08977 [stat.CO].arXiv
  54. [54]Vahid Satarifard et al. High-Fidelity Tuning of Olfactory Mixture Distances in the Perceptual Space of Smell Through a Community Effort. Dec. 2025.DOI
  55. [55]W. James Murdoch et al. “Definitions, Methods, and Applications in Interpretable Machine Learning”. In: Proceedings of the National Academy of Sciences 116.44 (Oct. 29, 2019), pp. 22071–22080.DOI
  56. [56]Templeton, A., & al., . et . (2024). Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet. Transformer Circuits Thread. <a href="https://doi.org/https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html" rel="noopener noreferrer" target="_blank">https://doi.org/https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html</a>
  57. [57]Joseph, S., & al., . et . (2025). Prisma: An Open Source Toolkit for Mechanistic Interpretability in Vision and Video. 2504.19475
  58. [58]Zvyagin, M., & al., . et . (2023). GenSLMs: Genome-scale Language Models Reveal SARS-CoV-2 Evolutionary Dynamics. The International Journal of High Performance Computing Applications, 37.6, 683–705.
  59. [59]Lipinski, C. A., & al., . et . (1997). Experimental and Computational Approaches to Estimate Solubility and Permeability in Drug Discovery and Development Settings. Advanced Drug Delivery Reviews, 23.1–3, 3–25.
  60. [60]Guillaume Alain and Yoshua Bengio. Understanding Intermediate Layers Using Linear Classifier Probes. Nov. 22, 2018. arXiv: 1610.01644 [stat]. Pre-published.arXiv
  61. [61]Shigeaki Kikuchi. “A History of the Structural Theory of Benzene - The Aromatic Sextet Rule and Huckel’s Rule”. In: Journal of Chemical Education 74.2 (Feb. 1, 1997), p. 194.DOI
  62. [62]Utkarsh Sharma and Jared Kaplan. A Neural Scaling Law from the Dimension of the Data Manifold. Apr. 22, 2020. arXiv: 2004.10802 [cs, stat]. Pre-published.arXiv
  63. [63]Nguyen, E., & al., . et . (2024). Sequence Modeling and Design from Molecular to Genome Scale with Evo. Science, 386.6723, eado9336.
  64. [64]You, Y., & al., . et . (2020). Large Batch Optimization for Deep Learning: Training BERT in 76 Minutes. 1904.00962
  65. [65]Alex Krizhevsky. One Weird Trick for Parallelizing Convolutional Neural Networks. Apr. 26, 2014. arXiv: 1404.5997 [cs]. Pre-published.arXiv
  66. [66]Bahri, Y., & al., . et . (2024). Explaining Neural Scaling Laws. Proceedings of the National Academy of Sciences, 121(27), e2311878121.
  67. [67]Sharma, U., & Kaplan, J. (2022). Scaling Laws from the Data Manifold Dimension. J. Mach. Learn. Res., 23(1), 9:343–9:376.
  68. [68]Anuj K. Nayak and Lav R. Varshney. “An Information Theory of Compute-Optimal Size Scaling, Emergence, and Plateaus in Language Models”. In: Workshop on Machine Learning and Compression, NeurIPS 2024. arXiv, Oct. 15, 2024.arxiv.org/abs/2410.01243
  69. [69]Ellis-Mohr, A. R., Nayak, A. K., & Varshney, L. R. (2025). A Theory of Inference Compute Scaling: Reasoning through Directed Stochastic Skill Search. Philosophical Transactions of the Royal Society A.
  70. [70]Muennighoff, N., & al., . et . (2023). Scaling Data-Constrained Language Models. http://arxiv.org/abs/2305.16264
  71. [71]Sorscher, B., & al., . et . (2023). Beyond Neural Scaling Laws: Beating Power Law Scaling via Data Pruning. http://arxiv.org/abs/2206.14486
  72. [72]Brown, T., & al., . et . (2020). Language Models Are Few-Shot Learners. Advances in Neural Information Processing Systems, 33, 1877–1901.
  73. [73]Huggingface/Transformers: Transformers: State-of-the-art Machine Learning for Pytorch, Tensor-Flow, and JAX. (2024). Huggingface/Transformers: Transformers: State-of-the-art Machine Learning for Pytorch, Tensor-Flow, and JAX.
  74. [74]Microsoft/DeepSpeed. (2024). Microsoft/DeepSpeed.
  75. [75]Falcon, W., & team, T. P. L. (2023). PyTorch Lightning. Zenodo.
  76. [76]Jerry Ma and Denis Yarats. On the Adequacy of Untuned Warmup for Adaptive Optimization. Mar. 19, 2021. arXiv: 1910.04209 [cs, stat]. URL: http://arxiv.org/abs/1910.04209. Pre-published.
  77. [77]Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. Jan. 4, 2019. arXiv: 1711.05101 [cs]. Pre-published.arXiv
  78. [78]Lightning-AI Torchmetrics. (2024). Lightning-AI Torchmetrics.
  79. [79]Jörg Degen et al. “On the Art of Compiling and Using ’Drug-Like’ Chemical Fragment Spaces”. In: ChemMedChem 3.10 (Oct. 20, 2008), pp. 1503–1507.DOI
  80. [80]Barbara Zdrazil et al. “The ChEMBL Database in 2023: A Drug Discovery Platform Spanning Multiple Bioactivity Data Types and Time Periods”. In: Nucleic Acids Research 52.D1 (Jan. 5, 2024), pp. D1180–D1192.DOI
  81. [81]John J. Irwin et al. “ZINC20—A Free Ultralarge-Scale Chemical Database for Ligand Discovery”. In: J. Chem. Inf. Model. 60.12 (Dec. 28, 2020), pp. 6065–6073.DOI
  82. [82]Kevin L Gering. “Prediction of electrolyte conductivity: results from a generalized molecular model based on ion solvation and a chemical physics framework”. In: Electrochimica Acta 225 (2017), pp. 175–189.DOI
  83. [83]Patrice Porion et al. “Comparative Study on Transport Properties for LiFAP and LiPF6 in Alkyl-Carbonates as Electrolytes through Conductivity, Viscosity and NMR Self-Diffusion Measurements”. In: Electrochimica Acta 114 (Dec. 30, 2013), pp. 95–104.DOI
  84. [84]Yumin Zhang, Imanuel Bier, and Venkatasubramanian Viswanathan. “Predicting Electrolyte Conductivity Directly from Molecular-Level Interactions”. In: ACS Energy Lett. 7.11 (Nov. 11, 2022), pp. 4061–4070.DOI
  85. [85]Sunwook Hwang et al. “Ionic conduction and solution structure in LiPF6 and LiBF4 propylene carbonate electrolytes”. In: The Journal of Physical Chemistry C 122.34 (2018), pp. 19438–19446.DOI
  86. [86]Zhenqin Wu et al. “MoleculeNet: A Benchmark for Molecular Machine Learning”. In: Chem. Sci. 9.2 (2018), pp. 513–530.DOI
  87. [87]Landrum, G. (2024). RDKit: Open-source Cheminformatics.
  88. [88]Alexandra Wahab et al. “The COMPAS Project : A Computational Database of Polycyclic Aromatic Systems . Phase 1: Cata - Condensed Polybenzenoid Hydrocarbons”. In: Journal of Chemical Information and Modeling 62.16 (Aug. 22, 2022), pp. 3704–3713.DOI
  89. [89]Eduardo Mayo Yanes, Sabyasachi Chakraborty, and Renana Gershoni-Poranne. “COMPAS-2 : A Dataset of Cata-Condensed Hetero-Polycyclic Aromatic Systems”. In: Scientific Data 11.1 (Jan. 19, 2024), p. 97.DOI
  90. [90]Alexandra Wahab and Renana Gershoni-Poranne. COMPAS-3 : A Data Set of Peri- Condensed Polybenzenoid Hydrocarbons. Feb. 26, 2024. URL: https://chemrxiv.org/engage/chemrxiv/article-details/65d8c60ae9ebbb4db90f6276. Pre-published.chemrxiv.org/engage/chemrxiv/articledetails/65d8c60ae9ebbb4db90f6276
  91. [91]L. J. Schaad and B. A. Jr. Hess. “Huckel theory and aromatically”. In: Journal of Chemical Education 51.10 (1974), p. 640. eprint: https://doi.org/10.1021/ed051p640.
  92. [92]Landrum, G., & al., . et . (2024). Rdkit/Rdkit: 2024 09 4 (Q3 2024) Release. In Zenodo.
  93. [93]Bezanson, J., & al., . et . (2017). Julia: A Fresh Approach to Numerical Computing. SIAM Review, 59(1), 65–98.
  94. [94]Papp, T. K., & al., . et . (2025). Tpapp/DynamicHMC.Jl: V3.5.0. In Zenodo.
  95. [95]Hoffman, M. D., & Gelman, A. (2014). The No-u-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research, 15(47), 1593–1623.
  96. [96]Aki Vehtari et al. “Rank-Normalization, Folding, and Localization: An Improved Rˆ for Assessing Convergence of MCMC (with Discussion)”. In: Bayesian Anal. 16.2 (June 1, 2021).DOI
  97. [97]Gelman, A., & al., . et . (2014). Bayesian Data Analysis. In Chapman and Hall/CRC (Number 3rd ed.).
  98. [98]Maziarka, L., & al., . et . (2020). Molecule Attention Transformer. 2002.08264
  99. [99]Juncai Li and Xiaofei Jiang. “Mol-BERT: An Effective Molecular Representation with BERT for Molecular Property Prediction”. In: Wireless Communications and Mobile Computing 2021.1 (Jan. 2021). Ed. by Yulin Wang, p. 7181815.DOI
  100. [100]Yüksel, A., & al., . et . (2023). SELFormer: Molecular Representation Learning via SELFIES Language Models. Mach. Learn.: Sci. Technol., 4(2), 025035.
  101. [101]Naifeng Wen et al. “A Fingerprints Based Molecular Property Prediction Method Using the BERT Model”. In: J Cheminform 14.1 (Oct. 21, 2022), p. 71.DOI
  102. [102]Ahmad, W., & al., . et . (2022). ChemBERTa-2: Towards Chemical Foundation Models. 2209.01712
  103. [103]Sultan, A., & al., . et . (2025). Transformers for Molecular Property Prediction: Lessons Learned from the Past Five Years. 2404.03969
  104. [104]Shion Honda, Shoi Shi, and Hiroki R. Ueda. SMILES Transformer: Pre-trained Molecular Fingerprint for Low Data Drug Discovery. Nov. 12, 2019. arXiv: 1911.04738 [cs, stat]. URL: http://arxiv.org/abs/1911.04738. Pre-published.
  105. [105]Jannis Born and Matteo Manica. “Regression Transformer Enables Concurrent Sequence Regression and Generation for Molecular Language Modelling”. In: Nature Machine Intelligence 5.4 (Apr. 2023), pp. 432–444.DOI
  106. [106]Ross Irwin et al. “Chemformer: A Pre-Trained Transformer for Computational Chemistry”. In: Mach. Learn.: Sci. Technol. 3.1 (Mar. 1, 2022), p. 015022.DOI
  107. [107]Dongyu Xue et al. “X-MOL: Large-Scale Pre-Training for Molecular Understanding and Diverse Molecular Analysis”. In: Science Bulletin 67.9 (May 2022), pp. 899–902.DOI
  108. [108]Rajaonson, E. M., & al., . et . (2025). CheMixHub: Datasets and Benchmarks for Chemical Mixture Property Prediction. 2506.12231
  109. [109]Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization. 2017. arXiv: 1412.6980 [cs.LG].arxiv.org/abs/1412.6980
  110. [110]Takase, S., & al., . et . (2024). Spike No More: Stabilizing the Pre-training of Large Language Models. https://doi.org/2312.16903
  111. [111]Taniguchi, S., & al., . et . (2024). ADOPT: Modified Adam Can Converge with Any B_{2} with the Optimal Rate. https://doi.org/2411.02853
  112. [112]Xiong, R., & al., . et . (2020). On Layer Normalization in the Transformer Architecture. Proceedings of Machine Learning Research, 119, 10524–10533.
  113. [113]Shoeybi, M., & al., . et . (2019). Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism. In CoRR: Abs/1909.08053. https://doi.org/1909.08053
  114. [114]David Weininger. “SMILES, a Chemical Language and Information System. 1. Introduction to Methodology and Encoding Rules”. In: J. Chem. Inf. Comput. Sci. 28.1 (Feb. 1, 1988), pp. 31–36.DOI
  115. [115]Mario Krenn et al. “Self-Referencing Embedded Strings (SELFIES): A 100% Robust Molecular String Representation”. In: Mach. Learn.: Sci. Technol. 1.4 (Dec. 1, 2020), p. 045024.DOI
  116. [116]Su, J., & al., . et . (2023). RoFormer: Enhanced Transformer with Rotary Position Embedding. https://doi.org/2104.09864
  117. [117]Krenn, M., & al., . et . (2022). SELFIES and the Future of Molecular String Representations. Patterns, 3(10), 100588.
  118. [118]Philippe Schwaller et al. ““Found in Translation”: Predicting Outcomes of Complex Organic Chemistry Reactions Using Neural Sequence-to-Sequence Models”. In: Chem. Sci. 9.28 (2018), pp. 6091–6098.DOI
  119. [119]James, C. A. (2016). OpenSMILES.
  120. [120]Goldman, O., & al., . et . (2024). Unpacking Tokenization: Evaluating Text Compression and Its Correlation with Model Performance. https://doi.org/2403.06265
  121. [121]Rust, P., & al., . et . (2021). How Good Is Your Tokenizer? On the Monolingual Performance of Language Models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) (pp. 3118–3135).
  122. [122]Thamme Gowda and Jonathan May. “Finding the Optimal Vocabulary Size for Neural Machine Translation”. In: Findings of the Association for Computational Linguistics: EMNLP 2020. Findings 2020. Ed. by Trevor Cohn, Yulan He, and Yang Liu. Online: Association for Computational Linguistics, Nov. 2020, pp. 3955–3964.DOI
  123. [123]Groom, C. R. et . al . (2016). The Cambridge Structural Database. Acta Crystallogr B Struct Sci Cryst Eng Mater, 72(2), 171–179.
  124. [124]Dilip Krishnamurthy et al. “Closed-Loop Electrolyte Design for Lithium-Mediated Ammonia Synthesis”. In: ACS Central Science 7.12 (2021), pp. 2073–2082. eprint: https://doi.org/10.1021/acscentsci.1c01151.
  125. [125]Laurence, C., & Gal, J. F. (2009). Lewis Basicity and Affinity Scales: Data and Measurement.
  126. [126]Pan Zhou, Yong Xiang, and Kai Liu. “Understanding and Applying the Donor Number of Electrolytes in Lithium Metal Batteries”. In: Energy Environ. Sci. 17.21 (2024), pp. 8057–8077.DOI
  127. [127]Chen, J. et . al . (2023). Design of Localized High-Concentration Electrolytes via Donor Number. ACS Energy Lett., 8(4), 1723–1734.
  128. [128]Jijian Xu et al. “Electrolyte Design for Li-ion Batteries under Extreme Operating Conditions”. In: Nature 614.7949 (Feb. 23, 2023), pp. 694–700.DOI
  129. [129]Sunghwan Kim et al. “PubChem 2025 Update”. In: Nucleic Acids Research 53.D1 (Jan. 6, 2025), pp. D1516–D1525.DOI
  130. [130]Théophile Gaudin, Patricia Rotureau, and Guillaume Fayet. “Mixture Descriptors toward the Development of Quantitative Structure–Property Relationship Models for the Flash Points of Organic Mixtures”. In: Industrial & Engineering Chemistry Research 54.25 (July 1, 2015), pp. 6596–6604.DOI
  131. [131]Dimroth, & Reichardt. (n.d.). Dimroth and Reichardt ET..
  132. [132]Leito, I. et . al . (2025). Acid Dissociation Constants in Selected Dipolar Non-Hydrogen-Bond-Donor Solvents (IUPAC Technical Report). Pure and Applied Chemistry, 97(9), 973–998.
  133. [133]Maho Nakata and Toshiyuki Maeda. “PubChemQC B3LYP/6-31G*//PM6 Data Set: The Electronic Structures of 86 Million Molecules Using B3LYP/6-31G* Calculations”. In: J. Chem. Inf. Model. 63.18 (Sept. 25, 2023), pp. 5734–5754.DOI
  134. [134]Maho Nakata and Tomomi Shimazaki. “PubChemQC Project: A Large-Scale First-Principles Electronic Structure Database for Data-Driven Chemistry”. In: J. Chem. Inf. Model. 57.6 (June 26, 2017), pp. 1300–1308.DOI
  135. [135]Ali Eftekhari and Dong-Won Kim. “Sodium-ion batteries: new opportunities beyond energy storage by lithium”. In: Journal of power sources 395 (2018), pp. 336–348.DOI
  136. [136]Lei, S. et . al . (2023). Fast-charging of lithium-ion batteries: A review of electrolyte design aspects. Battery Energy, 2(5), 20230018.
  137. [137]Qian Dong et al. “ILThermo: A Free-Access Web Database for Thermodynamic Properties of Ionic Liquids”. In: J. Chem. Eng. Data 52.4 (July 1, 2007), pp. 1151–1159.DOI
  138. [138]Romolo Francesconi, Carlo Castellari, and Fabio Comelli. “Excess Molar Enthalpies and Excess Molar Volumes of Diethyl Carbonate + Some N-Alkoxyethanols at (298.15 and 313.15) K”. In: J. Chem. Eng. Data 44.6 (Nov. 1, 1999), pp. 1373–1378.DOI
  139. [139]Fabio Comelli, Stefano Ottani, and Romolo Francesconi. “Excess Molar Enthalpies and Excess Molar Volumes of Dimethyl Carbonate + Seven Alkyl Acetates at 298.15 K”. In: J. Chem. Eng. Data 42.6 (Nov. 1, 1997), pp. 1208–1211.DOI
  140. [140]Lilei Zhang et al. “Density, Speed of Sound, Viscosity, Excess Properties, and Intermolecular Interaction of Methyl Tert-Butyl Ether (1) + Isopropanol (2) at 0.1 MPa and (293.15–313.15) K”. In: J. Chem. Eng. Data (June 12, 2025).DOI
  141. [141]Chunping Li et al. “Density, Viscosity, and Excess Properties for 1,2-Diaminoethane + 1,2-Ethanediol at (298.15, 303.15, and 308.15) K”. In: J. Chem. Eng. Data 55.9 (Sept. 9, 2010), pp. 4104–4107.DOI
  142. [142]Romolo Francesconi and Fabio Comelli. “Excess Molar Enthalpies, Densities, and Excess Molar Volumes of Binary Mixtures Containing Esters of Carbonic Acid at 298.15 and 313.15 K”. In: J. Chem. Eng. Data 40.4 (July 1, 1995), pp. 811–814.DOI
  143. [143]Fabio Comelli and Romolo Francesconi. “Excess Molar Enthalpies and Excess Molar Volumes of Propionic Acid + Octane, + Cyclohexane, + 1,3,5-Trimethylbenzene, + Oxane, or + 1,4-Dioxane at 313.15 K”. In: J. Chem. Eng. Data 41.1 (Jan. 1, 1996), pp. 101–104.DOI
  144. [144]Lukasz Ruszczyński et al. “Excess Enthalpies in Binary Systems of Isomeric C8 Aliphatic Monoethers with Acetonitrile and Their Description by the COSMO-SAC Model”. In: J. Chem. Eng. Data 61.2 (Feb. 11, 2016), pp. 996–1002.DOI
  145. [145]Jui-Tang Chen and Wong-Chao Chang. “Densities and Viscosities of Binary Mixtures of Propylene Glycol Monomethyl Ether Acetate with Methacrylic Acid, Benzyl Methacrylate, and 2-Hydroxyethyl Methacrylate between 298.15 K and 318.15 K”. In: J. Chem. Eng. Data 50.5 (Sept. 1, 2005), pp. 1753–1756.DOI
  146. [146]Abdullah S. Alkindi, Yahya M. Al-Wahaibi, and Ann H. Muggeridge. “Physical Properties (Density, Excess Molar Volume, Viscosity, Surface Tension, and Refractive Index) of Ethanol + Glycerol”. In: J. Chem. Eng. Data 53.12 (Dec. 11, 2008), pp. 2793–2796.DOI
  147. [147]Amalendu Pal and Sanjay Sharma. “Excess Molar Volumes and Viscosities of 1-Propanol + Ethylene Glycol, + Ethylene Glycol Monomethyl, + Ethylene Glycol Dimethyl, + Diethylene Glycol Dimethyl, + Triethylene Glycol Dimethyl, + Diethylene Glycol Diethyl, and + Diethylene Glycol Dibutyl Ethers at 298.15 K”. In: J. Chem. Eng. Data 43.4 (July 1, 1998), pp. 532–536.DOI
  148. [148]Ayyaz Muhammad et al. “Density and Excess Properties of Aqueous N-Methyldiethanolamine Solutions from (298.15 to 338.15) K”. In: J. Chem. Eng. Data 53.9 (Sept. 1, 2008), pp. 2217–2221.DOI
  149. [149]Qingguo Zhang et al. “Density, Electrical Conductivity, Dynamic Viscosity, Excess Properties, and Molecular Interactions of Ionic Liquid 1-Cyanopropyl-3-methylimidazolium Tetrafluoroborate and Binary System with Acetonitrile”. In: Journal of Chemical & Engineering Data 63.5 (Apr. 2018), pp. 1256–1265.DOI
  150. [150]Amin Pourmohammadbagher and John M. Shaw. “Excess Enthalpy and Excess Volume for Pyridine + Methyldiethanolamine and Pyridine + Ethanolamine Mixtures”. In: J. Chem. Eng. Data 58.8 (Aug. 8, 2013), pp. 2202–2209.DOI
  151. [151]Romolo Francesconi and Fabio Comelli. “Excess Enthalpies and Excess Volumes of the Liquid Binary Mixtures of Propylene Carbonate + Six Alkanols at 298.15 K”. In: Journal of Chemical & Engineering Data 41.6 (Jan. 1996), pp. 1397–1400.DOI
  152. [152]G. Hahn and P. Svejda. “Excess Volumes and Excess Enthalpies of Binary Liquid Mixtures of Trans-1,2-Dichloroethene and of Cis-1,2-Dichloroethene + n-Alkanes or +2,2,4-Trimethylpentane”. In: J. Chem. Eng. Data 41.3 (Jan. 1, 1996), pp. 619–623.DOI
  153. [153]Stefano Ottani, Fabio Comelli, and Carlo Castellari. “Densities, Viscosities, and Excess Molar Enthalpies of Propylene Carbonate + Anisole or + PhenetoIe at (293.15, 303.15, and 313.15) K”. In: Journal of Chemical & Engineering Data 46.1 (Dec. 2000), pp. 125–129.DOI
  154. [154]Xiaoqin Shi et al. “Density, Viscosity, and Excess Properties of Binary Mixtures of 2-(Methylamino)Ethanol with 2-Methoxyethanol, 2-Ethoxyethanol, and 2-Butoxyethanol from 293.15 to 353.15 K”. In: J. Chem. Eng. Data 64.9 (Sept. 12, 2019), pp. 3960–3970.DOI
  155. [155]Alain Valtz, Frédérick de Meyer, and Christophe Coquelet. “Density, Viscosity, and Excess Properties of Aqueous Solution of 1,3-Dimethyl-2-imidazolidinone (DMI)”. In: J. Chem. Eng. Data 68.4 (Apr. 13, 2023), pp. 781–792.DOI
  156. [156]Ren Ren et al. “Density, Excess Molar Volume and Conductivity of Binary Mixtures of the Ionic Liquid 1,2-Dimethyl-3-hexylimidazolium Bis(Trifluoromethylsulfonyl)Imide and Dimethyl Carbonate”. In: J. Chem. Eng. Data 56.1 (Jan. 13, 2011), pp. 27–30.DOI
  157. [157]Yingying Zuo, Xiguang Chen, and Jing Tong. “Excess Properties of Tetrabutylammonium Bromide-Acetic Acid Deep Eutectic Solvent Capable of High-Efficiency Denitration”. In: J. Chem. Eng. Data 69.1 (Jan. 11, 2024), pp. 102–113.DOI
  158. [158]Hsu-Chen Ku and Chein-Hsiun Tu. “Density and Viscosity of Binary Mixtures of Propan-2-ol, 1-Chlorobutane, and Acetonitrile”. In: J. Chem. Eng. Data 43.3 (May 1, 1998), pp. 465–468.DOI
  159. [159]Gangfeng Ouyang et al. “Excess Molar Volumes and Surface Tensions of Xylene with Isopropyl Ether or Methyl Tert-Butyl Ether at 298.15 K”. In: J. Chem. Eng. Data 49.3 (May 1, 2004), pp. 732–734.DOI
  160. [160]Cristina Vallés et al. “Excess Enthalpy, Density, Viscosity, and Speed of Sound for the Mixture Tetrahydropyran + 1-Butanol at (283.15, 298.15, and 313.15) K”. In: J. Chem. Eng. Data 49.5 (Sept. 1, 2004), pp. 1460–1464.DOI
  161. [161]Pragati Jain and Mukhtar Singh. “Density, Viscosity, and Excess Properties of Binary Liquid Mixtures of Propylene Carbonate with Polar and Nonpolar Solvents”. In: J. Chem. Eng. Data 49.5 (Sept. 1, 2004), pp. 1214–1217.DOI
  162. [162]Fuchun Chen et al. “Density, Viscosity, Speed of Sound, Excess Property and Bulk Modulus of Binary Mixtures of γ-Butyrolactone with Acetonitrile, Dimethyl Carbonate, and Tetrahydrofuran at Temperatures (293.15 to 333.15) K”. In: Journal of Molecular Liquids 209 (Sept. 2015), pp. 683–692.DOI
  163. [163]Malyanah Mohd Taib and Thanapalan Murugesan. “Density, Refractive Index, and Excess Properties of 1-Butyl-3-methylimidazolium Tetrafluoroborate with Water and Monoethanolamine”. In: J. Chem. Eng. Data 57.1 (Jan. 12, 2012), pp. 120–126.DOI
  164. [164]Ana M. Mainar et al. “Excess Molar Enthalpies of Mixtures of (±)-Linalool with Several Alkanols”. In: J. Chem. Eng. Data 56.10 (Oct. 13, 2011), pp. 3818–3822.DOI
  165. [165]Lugo, L., & al., . et . (2002). (p, Vm, T, x) measurements of dimethyl carbonate + octane binary mixtures. Fluid Phase Equilibria, 199(1–2), 135–145.
  166. [166]Francisca Maria Rodrigues Mesquita et al. “Density, Viscosities, and Excess Properties for Binary Mixtures of Sulfolane + Alcohols and Sulfolane + Glycols at Different Temperatures”. In: J. Chem. Eng. Data 59.7 (July 10, 2014), pp. 2196–2206.DOI
  167. [167]Gyan Prakash Dubey and Krishan Kumar. “Density, Speed of Sound, Viscosity, Excess Properties, and Prigogine–Flory–Patterson (PFP) Theory of Binary Mixtures of Amine and Alcohols”. In: J. Chem. Eng. Data 61.6 (June 9, 2016), pp. 1967–1980.DOI
  168. [168]Cristina Vallés et al. “Excess Enthalpy, Density, Speed of Sound, and Viscosity for 2-Methyltetrahydrofuran + 1-Butanol at (283.15, 298.15, and, 313.15) K”. In: J. Chem. Eng. Data 51.3 (May 11, 2006), pp. 1105–1109.DOI
  169. [169]Trevor M. Letcher and Gan G. Redhi. “Excess Enthalpies and Volumes for Mixtures of (Acetonitrile + a Carboxylic Acid) at 298.15 K”. In: Journal of Chemical & Engineering Data 45.1 (Nov. 1999), pp. 57–60.DOI
  170. [170]Zhao, Y., & al., . et . (2000). Effect of Temperature on Excess Molar Volumes and Viscosities for Propylene Carbonate + N,N-Dimethylformamide Mixtures. Journal of Chemical & Engineering Data, 45(3), 440–444.
  171. [171]Amalendu Pal and Harsh Kumar. “Excess Molar Volumes and Viscosities of Mixtures Containing Some Polyethers + Acetonitrile at 298.15 K”. In: J. Chem. Eng. Data 44.6 (Nov. 1, 1999), pp. 1330–1334.DOI
  172. [172]René A. Clará et al. “Density, Viscosity, Vapor-Liquid Equilibrium, and Excess Molar Enthalpy of [Chloroform + Methyl Tert-Butyl Ether]”. In: J. Chem. Eng. Data 55.12 (Dec. 9, 2010), pp. 5862–5867.DOI
  173. [173]Elena N. Rezanova, Klaus Kammerer, and Rüdiger N. Lichtenthaler. “Excess Properties of Binary Alkanol + Diisopropyl Ether (DIPE) or + Dibutyl Ether (DBE) Mixtures and the Application of the Extended Real Associated Solution Model”. In: J. Chem. Eng. Data 44.6 (Nov. 1, 1999), pp. 1235–1239.DOI
  174. [174]Oana Ciocirlan and Olga Iulian. “Properties of Pure 1-Butyl-2,3-Dimethylimidazolium Tetrafluoroborate Ionic Liquid and Its Binary Mixtures with Dimethyl Sulfoxide and Acetonitrile”. In: J. Chem. Eng. Data 57.11 (Nov. 8, 2012), pp. 3142–3148.DOI
  175. [175]Wei Fan et al. “Density, Excess Molar Volume, and Viscosity for the Methyl Methacrylate + 1-Butyl-3-methylimidazolium Hexafluorophosphate Ionic Liquid Binary System at Atmospheric Pressure”. In: Journal of Chemical & Engineering Data 54.8 (May 2009), pp. 2307–2311.DOI
  176. [176]Changsheng Yang, Wei Xu, and Peisheng Ma. “Excess Molar Volumes and Viscosities of Binary Mixtures of Dimethyl Carbonate with Chlorobenzene, Hexane, and Heptane from (293.15 to 353.15) K and at Atmospheric Pressure”. In: Journal of Chemical & Engineering Data 49.6 (2004), pp. 1802–1808. eprint: https://doi.org/10.1021/je049777o.
  177. [177]Fabio Comelli, Romolo Francesconi, and Stefano Ottani. “Excess Molar Enthalpies of Binary Mixtures Containing Propylene Carbonate + 23 Alkanoates at 298.15 K”. In: Journal of Chemical & Engineering Data 43.3 (Mar. 1998), pp. 333–336.DOI
  178. [178]Comelli, F. (2006). Excess Molar Enthalpies, Molar Heat Capacities, Densities, Viscosities, and Refractive Indices of Dimethyl Sulfoxide + Esters of Carbonic Acid at 308.15 K and Atmospheric Pressure. Journal of Chemical & Engineering Data, 51(2), 665–670.
  179. [179]Muhuri, P. K., Das, B., & Hazra, D. K. (1996). Viscosities and Excess Molar Volumes of Binary Mixtures of Propylene Carbonate with Tetrahydrofuran and Methanol at Different Temperatures. Journal of Chemical & Engineering Data, 41(6), 1473–1476.
  180. [180]Reza Tahery, Hamid Modarress, and John Satherley. “Density and Surface Tension of Binary Mixtures of Acetonitrile + 1-Alkanol at 293.15 K”. In: J. Chem. Eng. Data 51.3 (May 11, 2006), pp. 1039–1042.DOI
  181. [181]Roy, M. N., Sinha, B., & Dakua, V. K. (2006). Excess Molar Volumes and Viscosity Deviations of Binary Liquid Mixtures of 1,3-Dioxolane and 1,4-Dioxane with Butyl Acetate, Butyric Acid, Butylamine, and 2-Butanone at 298.15 K. Journal of Chemical & Engineering Data, 51(2), 590–594.
  182. [182]Francesconi, R., & Comelli, F. (1994). Excess molar volumes of binary mixtures containing dimethyl carbonate + linear and cyclic ethers. Journal of Chemical & Engineering Data, 39(1), 106–107.
  183. [183]Yamamoto, H. (1997). Density and Excess Molar Volume of Tri-n-Octylamine + Propionic Acid + Diluent at 298.15 K. J. Chem. Eng. Data, 42(2), 238–242.
  184. [184]Álvarez, E. (2006). Density, Viscosity, Excess Molar Volume, and Viscosity Deviation of Three Amyl Alcohols + Ethanol Binary Mixtures from 293.15 to 323.15 K. J. Chem. Eng. Data, 51(3), 940–945.
  185. [185]Wang, E. (2024). Excess Properties and Intermolecular Interaction of Three Butanol Isomers + 1,2-Propanediamine Binary Systems. J. Chem. Eng. Data, 69(3), 785–807.
  186. [186]Anwar, N., & Riyazuddeen. (2018). Excess Molar Volumes, Excess Molar Isentropic Compressibilities, Viscosity Deviations, and Activation Parameters for 1-Ethyl-3-methyl-imidazolium Trifluoromethanesulfonate + Dimethyl Sulfoxide and/or Acetonitrile at $T$ = 298.15 to 323.15 K and $P$ = 0.1 MPa. J. Chem. Eng. Data, 63(2), 269–289.
  187. [187]Lu, H. (2001). Excess Molar Volumes and Viscosities for Binary Mixtures of γ-Butyrolactone with Methyl Formate, Ethyl Formate, Methyl Acetate, Ethyl Acetate, and Acetonitrile at 298.15 K. J. Chem. Eng. Data, 46(3), 631–634.
  188. [188]Zaitseva, A. (2016). Vapor–Liquid Equilibria, Excess Enthalpy, and Density of Aqueous γ-Valerolactone Solutions. Journal of Chemical & Engineering Data, 61(2), 881–890.
  189. [189]Shafaat, S., & Almasi, M. (2017). Influence of Temperature and Carbon Chain on Thermophysical Properties of Benzaldehyde/Alkan-2-ol Binary Mixtures. Journal of Chemical & Engineering Data, 62(8), 2406–2412.
  190. [190]Barsainyan, A. A. (2023). OpenPOM - Open Principal Odor Map. <a href="https://doi.org/https://github.com/BioMachineLearning/openpom" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/BioMachineLearning/openpom</a>
  191. [192]Bradley, M. (2024). Radioactive Isotope Half Lives. From the Wolfram Data Repository.
  192. [193]Flajolet, P., & al., . et . (2007). HyperLogLog: The Analysis of a near-Optimal Cardinality Estimation Algorithm. Discrete Mathematics & Theoretical Computer Science DMTCS Proceedings, AH.
  193. [194]Hernández-Lobato, J. M., & al., . et . (2017). Parallel and Distributed Thompson Sampling for Large-Scale Accelerated Exploration of Chemical Space. Proceedings of the 34th International Conference on Machine Learning - Volume 70, 70, 1470–1479.
  194. [195]Moussa, J. E., & Stewart, J. J. P. (2025). MOPAC. Version v23.1.2. In Zenodo.
  195. [196]Frisch, M. J., & al., . et . (2009). Gaussian 09. In Gaussian, Inc..
  196. [197]Wang, S., & al., . et . (2020). Improving Conformer Generation for Small Rings and Macrocycles Based on Distance Geometry and Experimental Torsional-Angle Preferences. J. Chem. Inf. Model., 60(4), 2044–2058.
  197. [198]O’Boyle, N. M., & al., . et . (2011). Open Babel: An Open Chemical Toolbox. J Cheminform, 3(1), 33.
  198. [199]Rappe, A. K., & al., . et . (1992). UFF, a Full Periodic Table Force Field for Molecular Mechanics and Molecular Dynamics Simulations. J. Am. Chem. Soc., 114(25), 10024–10035.
  199. [200]Tosco, P., Stiefl, N., & Landrum, G. (2014). Bringing the MMFF Force Field to the RDKit: Implementation and Validation. J Cheminform, 6(1), 37.
  200. [201]Halgren, T. A. (1996). Merck Molecular Force Field. I. Basis, Form, Scope, Parameterization, and Performance of MMFF94. J. Comput. Chem., 17(5–6), 490–519.
  201. [202]Zhang, Y., & Viswanathan, V. (2020). Not All Fluorination Is the Same: Unique Effects of Fluorine Functionalization of Ethylene Carbonate for Tuning Solid-Electrolyte Interphase in Li Metal Batteries. Langmuir, 36(39), 11450–11466.
  202. [203]Shkrob, I. A., Wishart, J. F., & Abraham, D. P. (2015). What makes fluoroethylene carbonate different?. The Journal of Physical Chemistry C, 119(27), 14954–14964.
  203. [204]Michan, A. L., & al., . et . (2016). Fluoroethylene carbonate and vinylene carbonate reduction: understanding lithium-ion battery electrolyte additives and solid electrolyte interphase formation. Chemistry of Materials, 28(22), 8149–8159.
  204. [205]Li, T., & al., . et . (2019). Fluorinated solid-electrolyte interphase in high-voltage lithium metal batteries. Joule, 3(11), 2647–2661.
  205. [206]Xia, L., & al., . et . (2021). Recent advances in non-aqueous liquid electrolytes containing fluorinated compounds for high energy density lithium-ion batteries. Energy Storage Materials, 38, 542–570.
  206. [207]Li, Z., & al., . et . (2023). Critical review of fluorinated electrolytes for high-performance lithium metal batteries. Advanced Functional Materials, 33(32), 2300502.
  207. [208]Wang, P., & al., . et . (2025). Fluorine-free electrolytes for sustainable lithium batteries: a review. Npj Materials Sustainability, 3(1), 32.
  208. [209]He, M., & al., . et . (2020). The intrinsic behavior of lithium fluoride in solid electrolyte interphases on lithium. Proceedings of the National Academy of Sciences, 117(1), 73–79.
  209. [210]Frisch, M. ea ., & al., . et . (2016). Gaussian 16. In Gaussian.
  210. [211]Becke, A. D. (1993). Density-functional thermochemistry. III. The role of exact exchange. The Journal of chemical physics, 98(7), 5648–5652.
  211. [212]Lee, C., Yang, W., & Parr, R. G. (1988). Development of the Colle-Salvetti correlation-energy formula into a functional of the electron density. Physical Review B, 37(2), 785.
  212. [213]Stephens, P. J., & al., . et . (1994). Ab initio calculation of vibrational absorption and circular dichroism spectra using density functional force fields. The Journal of physical chemistry, 98(45), 11623–11627.
  213. [214]Weigend, F., & Ahlrichs, R. (2005). Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for H to Rn: Design and assessment of accuracy. Physical Chemistry Chemical Physics, 7(18), 3297–3305.
  214. [215]Grimme, S., & al., . et . (2010). A consistent and accurate ab initio parametrization of density functional dispersion correction (DFT-D) for the 94 elements H-Pu. The Journal of chemical physics, 132(15).
  215. [216]Reed, A. E., Weinstock, R. B., & Weinhold, F. (1985). Natural population analysis. The Journal of chemical physics, 83(2), 735–746.
  216. [217]Glendening, E. D., Landis, C. R., & Weinhold, F. (2013). NBO 6.0: Natural bond orbital analysis program. Journal of Computational Chemistry, 34(16), 1429–1437.
  217. [218]Varshney, L. R. (2019). Mathematical Limit Theorems for Computational Creativity. IBM Journal of Research and Development, 63(1), 2:1–2:12.
  218. [219]Varshney, L. R., Rajani, N. F., & Socher, R. (2020). Explaining Creative Artifacts. Proceedings of the ICML 2020 Workshop on Human Interpretability in Machine Learning.
  219. [220]Hero, A. O., & al., . et . (2002). Applications of Entropic Spanning Graphs. IEEE Signal Process. Mag., 19(5), 85–95.
  220. [221]Khetan, A., Pitsch, H., & Viswanathan, V. (2014). Identifying descriptors for solvent stability in nonaqueous Li–O2 batteries. The Journal of Physical Chemistry Letters, 5(8), 1318–1323.
  221. [222]Khetan, A., Luntz, A., & Viswanathan, V. (2015). Trade-offs in capacity and rechargeability in nonaqueous Li–O2 batteries: solution-driven growth versus nucleophilic stability. The Journal of Physical Chemistry Letters, 6(7), 1254–1259.
  222. [223]Aetukuri, N. B., & al., . et . (2015). Solvating additives drive solution-mediated electrochemistry and enhance toroid growth in non-aqueous Li–O2 batteries. Nature Chemistry, 7(1), 50–56.
  223. [224]McCloskey, B. D., & al., . et . (2012). Limitations in rechargeability of Li-O2 batteries and possible origins. The Journal of Physical Chemistry Letters, 3(20), 3043–3047.
  224. [225]Luntz, A. C., & al., . et . (2013). Tunneling and polaron charge transport through Li2O2 in Li–O2 batteries. The Journal of Physical Chemistry Letters, 4(20), 3494–3499.
  225. [226]Khetan, A., Pitsch, H., & Viswanathan, V. (2014). Solvent degradation in nonaqueous Li-O2 batteries: oxidative stability versus H-abstraction. The Journal of Physical Chemistry Letters, 5(14), 2419–2424.
  226. [227]Zhang, S. S., & Read, J. (2011). Partially fluorinated solvent as a co-solvent for the non-aqueous electrolyte of Li/air battery. Journal of Power Sources, 196(5), 2867–2870.
  227. [228]Wijaya, O., & al., . et . (2015). A gamma fluorinated ether as an additive for enhanced oxygen activity in Li–O2 batteries. Journal of Materials Chemistry A, 3(37), 19061–19067.
  228. [229]Burke, C. M., & al., . et . (2015). Enhancing electrochemical intermediate solvation through electrolyte anion selection to increase nonaqueous Li–O2 battery capacity. Proceedings of the National Academy of Sciences, 112(30), 9293–9298.
  229. [230]Zhao, Q., & al., . et . (2018). Binary mixtures of highly concentrated tetraglyme and hydrofluoroether as a stable and nonflammable electrolyte for Li–O2 batteries. ACS Applied Materials & Interfaces, 10(31), 26312–26319.
  230. [231]Schwenke, K. U., & al., . et . (2013). Stability of superoxide radicals in glyme solvents for non-aqueous Li–O2 battery electrolytes. Physical Chemistry Chemical Physics, 15(28), 11830–11839.
  231. [232]Kwak, W.-J., & al., . et . (2020). Optimized electrolyte with high electrochemical stability and oxygen solubility for lithium–oxygen and lithium–air batteries. ACS Energy Letters, 5(7), 2182–2190.
  232. [233]Zaidi, S. S. H., Douglas, J. T., & Li, X. (2024). Fluorobenzene-Diluted Localized Highly Concentrated Electrolyte for Enhanced Electrochemical Performance of Li–O2 Battery. ACS Applied Energy Materials, 7(19), 8610–8621.
  233. [234]Jiang, Z., & al., . et . (2022). Quenching singlet oxygen via intersystem crossing for a stable Li-O2 battery. Proceedings of the National Academy of Sciences, 119(34), e2202835119.
  234. [235]Petit, Y. K., & al., . et . (2019). DABCOnium: An Efficient and High-Voltage Stable Singlet Oxygen Quencher for Metal–O2 Cells. Angewandte Chemie International Edition, 58(20), 6535–6539.
  235. [236]Chen, J., & al., . et . (2022). Building a stable artificial solid electrolyte interphase on lithium metal anodes toward long-life Li–O2 batteries. Journal of Power Sources, 540, 231603.
  236. [237]Wen, Z., & al., . et . (2025). Boosting the Li-O2 pouch cell beyond 860 Wh kg-1 with an O2-enriched localized high-concentration electrolyte. National Science Review, 12(7), nwaf059.
  237. [238]Rigoni, A. S., Breedon, M., & Spencer, M. J. (2024). Use of Perfluorochemicals in Li-Air Batteries: A Critical Review. ACS Applied Materials & Interfaces, 16(21), 26967–26983.
  238. [239]Nasybulin, E., & al., . et . (2013). Effects of electrolyte salts on the performance of Li–O2 batteries. The Journal of Physical Chemistry C, 117(6), 2635–2645.
  239. [240]Ertl, P., & Schuffenhauer, A. (2009). Estimation of Synthetic Accessibility Score of Drug-like Molecules Based on Molecular Complexity and Fragment Contributions. Journal of Cheminformatics, 1(1), 8.
  240. [241]Coley, C. W., & al., . et . (2018). SCScore: Synthetic Complexity Learned from a Reaction Corpus. J. Chem. Inf. Model., 58(2), 252–261.
  241. [242]Sheridan, R. P., & al., . et . (2014). Modeling a Crowdsourced Definition of Molecular Complexity. J. Chem. Inf. Model., 54(6), 1604–1616.
  242. [243]Marshall, S. M., & al., . et . (2021). Identifying Molecules as Biosignatures with Assembly Theory and Mass Spectrometry. Nat Commun, 12(1).
  243. [244]Chen, S., & Jung, Y. (2024). Estimating the Synthetic Accessibility of Molecules with Building Block and Reaction-Aware SAScore. Journal of Cheminformatics, 16(1), 83.
  244. [245]Voršilák, M., & al., . et . (2020). SYBA: Bayesian Estimation of Synthetic Accessibility of Organic Compounds. J Cheminform, 12(1), 35.
  245. [246]Sterling, T., & Irwin, J. J. (2015). ZINC 15 – Ligand Discovery for Everyone. J. Chem. Inf. Model., 55(11), 2324–2337.
  246. [247]Voršilák, M., & Svozil, D. (2017). Nonpher: Computational Method for Design of Hard-to-Synthesize Structures. J Cheminform, 9(1), 20.
  247. [248]Seet, I., & al., . et . (2024). Rapid Computation of the Assembly Index of Molecular Graphs. https://doi.org/2410.09100
  248. [249]Choi, C., & al., . et . (2023). Understanding the Molecular Mechanisms of Odorant Binding and Activation of the Human OR52 Family. Nature Communications, 14(1), 8105.
  249. [250]Zarzo, M. (2012). Effect of Functional Group and Carbon Chain Length on the Odor Detection Threshold of Aliphatic Compounds. Sensors (Basel, Switzerland), 12(4), 4105–4112.
  250. [251]Genva, M., & al., . et . (2019). Is It Possible to Predict the Odor of a Molecule on the Basis of Its Structure?. International Journal of Molecular Sciences, 20(12), 3018.
  251. [252]Sagar, V., & al., . et . (2023). High-Precision Mapping Reveals the Structure of Odor Coding in the Human Brain. Nature Neuroscience, 26(9), 1595–1602.
  252. [253]Zhang, H., & al., . et . (2023). Hippocampal Spatial Representations Exhibit a Hyperbolic Geometry That Expands with Experience. Nature Neuroscience, 26(1), 131–139.
  253. [254]Decay Modes. (2021). Decay Modes. In Nuclear and Radiochemistry (pp. 119–186). <a href="https://doi.org/https://onlinelibrary.wiley.com/doi/pdf/10.1002/9783527831944.ch6" rel="noopener noreferrer" target="_blank">https://doi.org/https://onlinelibrary.wiley.com/doi/pdf/10.1002/9783527831944.ch6</a>
  254. [255]Weizsäcker, C. F. v . (1935). Zur Theorie der Kernmassen. Zeitschrift für Physik, 96(7), 431–458.
  255. [256]Dave, A., & al., . et . (2022). Autonomous Optimization of Non-Aqueous Li-ion Battery Electrolytes via Robotic Experimentation and Machine Learning Coupling. Nat Commun, 13(1), 5454.
  256. [257]De, S., & al., . et . (2017). Mapping and Classifying Molecules from a High-Throughput Structural Database. Journal of Cheminformatics, 9(1), 6.
  257. [258]Bartók, A. P., Kondor, R., & Csányi, G. (2013). On representing chemical environments. Physical Review B, 87(18).
  258. [259]Himanen, L., & al., . et . (2020). DScribe: Library of Descriptors for Machine Learning in Materials Science. Computer Physics Communications, 247, 106949.
  259. [260]Ikeda, M., & Aniya, M. (2013). Understanding the Vogel–Fulcher–Tammann Law in Terms of the Bond Strength–Coordination Number Fluctuation Model. Journal of Non-Crystalline Solids, 371–372, 53–57.
  260. [261]Dahbi, M., & al., . et . (2011). Comparative Study of EC/DMC LiTFSI and LiPF6 Electrolytes for Electrochemical Storage. Journal of Power Sources, 196(22), 9743–9750.
  261. [262]Rushing, J., & al., . et . (2022). Tale of a “Non-interacting” Additive in a Lithium-Ion Electrolyte: Effect on Ionic Speciation and Electrochemical Properties. The Journal of Physical Chemistry. C, Nanomaterials and Interfaces, 126(4), 2141–2150.
  262. [263]Berhaut, C. L., & al., . et . (2019). Ionic Association Analysis of LiTDI, LiFSI and LiPF6 in EC/DMC for Better Li-ion Battery Performances. RSC Advances, 9(8), 4599–4608.
  263. [264]Galle Kankanamge, S. R., & Kuroda, D. G. (2020). Molecular Structure, Chemical Exchange, and Conductivity Mechanism of High Concentration LiTFSI Electrolytes. The Journal of Physical Chemistry B, 124(10), 1965–1977.
  264. [265]Hall, D. S., Self, J., & Dahn, J. R. (2015). Dielectric Constants for Quantum Chemistry and Li-Ion Batteries: Solvent Blends of Ethylene Carbonate and Ethyl Methyl Carbonate. The Journal of Physical Chemistry C, 119(39), 22322–22330.
  265. [266]Lee, J., & al., . et . (2019). Tuning Two Interfaces with Fluoroethylene Carbonate Electrolytes for High-Performance Li/LCO Batteries. ACS Omega, 4(2), 3220–3227.
  266. [267]Thaler, S., & al., . et . (2024). Active Learning Graph Neural Networks for Partial Charge Prediction of Metal-Organic Frameworks via Dropout Monte Carlo. Npj Comput Mater, 10(1), 86.
  267. [268]Chung, Y. G., & al., . et . (2019). Advances, Updates, and Analytics for the Computation-Ready, Experimental Metal–Organic Framework Database: CoRE MOF 2019. J. Chem. Eng. Data, 64(12), 5985–5998.
  268. [269]Kancharlapalli, S., & al., . et . (2021). Fast and Accurate Machine Learning Strategy for Calculating Partial Atomic Charges in Metal–Organic Frameworks. J. Chem. Theory Comput., 17(5), 3052–3064.
  269. [270]Lehner, M. T., & al., . et . (2023). DASH: Dynamic Attention-Based Substructure Hierarchy for Partial Charge Assignment. J. Chem. Inf. Model., 63(19), 6014–6028.
  270. [271]Gurnee, W., & al., . et . (2023). Finding Neurons in a Haystack: Case Studies with Sparse Probing. 2305.01610
  271. [272]Elhage, N., & al., . et . (2022). Toy Models of Superposition. http://arxiv.org/abs/2209.10652
  272. [273]Lindsey, J., & al., . et . (2024). Sparse Crosscoders for Cross-Layer Features and Model Diffing. <a href="https://doi.org/https://transformer-circuits.pub/2024/crosscoders/index.html" rel="noopener noreferrer" target="_blank">https://doi.org/https://transformer-circuits.pub/2024/crosscoders/index.html</a>
  273. [274]Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning. In Springer Series in Statistics.
  274. [275]Leitao, J. C., & al., . et . (2016). Is This Scaling Nonlinear?. R. Soc. Open Sci., 3(7), 150649. 1604.02872
  275. [276]Wade, L. G., & Simek, J. (2016). Organic Chemistry.
  276. [277]Reichardt, C., & Welton, T. (2003). Solvents and Solvent Effects in Organic Chemistry (Vol. 3rd).
  277. [278]Grimme, S. (2004). Accurate description of van der Waals complexes by density functional theory including empirical corrections. Journal of Computational Chemistry, 25(12), 1463–1473.
  278. [279]Moses, W. S., & al., . et . (2021). Reverse-Mode Automatic Differentiation and Optimization of GPU Kernels via Enzyme. Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’21, 1–16.
  279. [280]Moses, W. S., & al., . et . (2022). Scalable Automatic Differentiation of Multiple Parallel Paradigms through Compiler Augmentation. Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis, SC ’22.
  280. [281]Jain, R., & Chlamtac, I. (1985). The P2 Algorithm for Dynamic Calculation of Quantiles and Histograms without Storing Observations. Commun. ACM, 28(10), 1076–1085.
  281. [282]Chung, Y., & al., . et . (2013). A Nondegenerate Penalized Likelihood Estimator for Variance Parameters in Multilevel Models. Psychometrika, 78(4), 685–709.
  282. [283]Gelman, A., & al., . et . (2025). Prior Choice Recommendations. Stan. <a href="https://doi.org/https://github.com/stan-dev/stan/wiki/Prior-Choice-Recommendations" rel="noopener noreferrer" target="_blank">https://doi.org/https://github.com/stan-dev/stan/wiki/Prior-Choice-Recommendations</a>
  283. [284]Piantadosi, S. T. (2014). Zipf’s Word Frequency Law in Natural Language: A Critical Review and Future Directions. Psychon Bull Rev, 21(5), 1112–1130.
  284. [285]Hernández-Fernández, A., & Torre, I. G. (2022). Compression Principle and Zipf’s Law of Brevity in Infochemical Communication. Biol. Lett., 18(7), 20220162.
  285. [286]Emin Martinian and Jonathan S. Yedidia. Iterative Quantization Using Codes On Graphs. Aug. 2, 2004. arXiv: cs/0408008. Pre-published.arXiv
  286. [287]Yin, M., & al., . et . (2024). Entropy Law: The Story Behind Data Compression and LLM Performance. 2407.06645
  287. [288]Kahan, W. (1965). Pracniques: Further Remarks on Reducing Truncation Errors. Commun. ACM, 8(1), 40.
  288. [289]Nowozin, S. (2016). Streaming Log-sum-exp Computation. Sebastian’s Slow Blog. <a href="https://doi.org/https://www.nowozin.net/sebastian/blog/streaming-log-sum-exp-computation.html" rel="noopener noreferrer" target="_blank">https://doi.org/https://www.nowozin.net/sebastian/blog/streaming-log-sum-exp-computation.html</a>
  289. [290]Papp, T. K., & contributors, S. jl . (2024). JuliaStats/LogExpFunctions.Jl.
  290. [291]Widmann, D., & Axen, S. (2025). TuringLang/MCMC​​DiagnosticTools.Jl.
  291. [292]BS EN 80000-13:2008: Quantities and Units: Information Science and Technology. (2008). BS EN 80000-13:2008: Quantities and Units: Information Science and Technology.
  292. [293]Martin, C. H., & al., . et . (2021). Predicting Trends in the Quality of State-of-the-Art Neural Networks without Access to Training or Testing Data. Nat Commun, 12(1), 4122. 2002.06716

Paper details

vlnc:W_437wg7xj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

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