WorksPhysics and Astronomy
PLUMED-JAX: GPU-native enhanced sampling for atomistic simulations
Abstract
We present PLUMED-JAX, an AI-assisted adaptation of selected PLUMED collective variables, umbrella sampling, and metadynamics for JAX-based atomistic simulations, developed using gpt-6.1-sol with reasoning effort set to high within the Codex harness. Integration with JAX-MD supports compiled molecular dynamics, batched replicas, walkers sharing a bias, and execution across multiple graphics processing units. We validate the framework against PLUMED and exact free energy benchmarks, and demonstrate applications to butane, silicon crystallization, and lithium transport, with severalfold improvements in computational throughput for both classical forcefields and machine learning interatomic potentials.
Introduction
Enhanced sampling methods enable the exploration of rare atomistic phenomena and their underlying free energy landscapes [1]. Recent applications illustrate the widespread impact of enhanced sampling such as across zeolite catalysis [2], ribosomal substrate selection [3], phase change materials for heat storage [4], and the physics of nanoconfined water [5]. In this context, the PLUMED library [6, 7, 8, 9] has played a central role in making enhanced sampling accessible and reproducible across forcefield and electronic structure codes. Notably, its collective variable library and modular design facilitate the development and testing of new sampling methods [10, 11, 12].
Atomistic simulations using classical forcefields and electronic structure methods have traditionally relied on software optimized for central processing units (CPUs), as have libraries such as PLUMED. Over the recent decades, graphics processing unit (GPU) acceleration has become increasingly established in forcefield [13, 14, 15] and electronic structure simulation codes [16, 17, 18]. The availability of GPUs has also supported the widespread adoption of machine learning interatomic potentials (MLIPs), which can approach the accuracy of their first-principles reference calculations at substantially lower computational cost.
GPU native simulation engines such as JAX-MD [19] and TorchSim [20] offer further acceleration of forcefield- and MLIP-based simulations through just-in-time (JIT) compilation and batched execution of multiple simulations. We note that these capabilities are attractive for enhanced sampling methods involving multiple replicas, which in principle can be trivially batched across a fixed number of simulation steps. Building on the precedent set by codes such as PySAGES [21], we therefore consider implementations of collective variables (CVs), bias forces, and bias updates that integrate with MLIPs and forcefields but also support batching and end-to-end JIT compilation to improve sampling efficiency and reduce execution time.
This work makes selected PLUMED functionality available within JAX and documents its validation against PLUMED. The resulting PLUMED-JAX library includes CVs and two enhanced sampling alalgorithms, umbrella sampling (US) [22] and (well-tempered) metadynamics (MTD) [23, 24], developed using gpt-6.1-sol(reasoning_effort=high) within the Codex harness. By using PLUMED-JAX with JAX-MD, we explore through three examples how batching and end-to-end compilation can support enhanced sampling with different physical potentials and computational requirements. For two lightweight forcefields, we batch simulations on a single GPU and combine this execution with replica exchange between umbrella windows and a shared MTD bias between walkers. Finally, an application with an MLIP illustrates how JAX’s sharding capabilities can be used to distribute independent umbrella windows across multiple GPUs while retaining batching within each device.
Implementation
Figure 1 summarizes the components of PLUMED-JAX, their interaction with the JAX-MD simulation driver, and the use of JAX utilities for parallel execution and compilation. We describe the core components, their implementation and design choices, and their validation against PLUMED.
Figure 1. Schematic of batched enhanced sampling with JAX-MD and PLUMED-JAX.
Collective variables
A collective variable is a function of the atomic coordinates that provides a low-dimensional description of a system’s microscopic state. Mathematically, a CV is simply a scalar function , where is the matrix containing the atomic positions as rows and is the matrix containing the three periodic lattice vectors when periodic boundary conditions are implied.
Each CV is implemented as a Python function using JAX operations, with derivatives with respect to atomic positions and the simulation cell obtained through JAX’s automatic differentiation [25]. Appendix § 7 describes the test configurations and validation procedure; Table A1 lists all CV types available in PLUMED-JAX, their scalar components, and the representative option combinations tested. The same table reports absolute differences from PLUMED in CV values, coordinate derivatives, and derivatives under simultaneous deformation of the coordinates and cell. The reference calculations use PLUMED 2.10.1 on a CPU, whereas PLUMED-JAX is evaluated on both a CPU and a GPU, using double-precision (64-bit floating-point) arithmetic.
Biasing potential
Enhanced sampling methods introduce a biasing potential that can depend on time. The total potential energy felt by the system is
where is the physical potential energy surface, is the biasing potential, is the vector of selected CVs, and is simulation time. The total force acting on atom is , where is the physical force and is the additional force from the bias.
Each bias is implemented as a Python function of the CVs (and time). JAX obtains the bias forces by differentiating this function composed with the CV functions, using backpropagation to apply the chain rule:
Thus, specifying the bias energy requires no separate implementation of its forces.
Umbrella sampling
US [22] typically uses a harmonic bias to keep the CV vector near a chosen center:
where labels an umbrella, is its center, and is a diagonal matrix of force constants for the CVs. For each periodic CV, the corresponding component of is replaced by the shortest signed difference between the two values.
Since each bias localizes sampling near its center, US typically uses multiple replicas with centers distributed across the full CV range of interest. Each replica is called an umbrella. An advantage of implementing US in JAX is that, when combined with a JAX implementation of the physical potential, as provided by JAX-MD in this work, independent umbrellas can be batched using jax.vmap when they fit on one GPU. Otherwise, they can be sharded across GPUs, with local batching using jax.vmap on each GPU.
Simultaneous simulation of the umbrellas also allows replica exchange, which can improve sampling by letting configurations move between neighboring umbrellas. This combination is known as replica exchange US (REUS) [26]. For umbrellas and at the same temperature, we propose the swap , where and contain the atom positions in the two replicas. Writing their CV vectors as and , the change in the sum of bias energies and the acceptance probability are
where is the inverse temperature. The physical energies cancel because all umbrellas use the same physical potential. On acceptance, positions and momenta are exchanged and the forces are updated for the receiving umbrella.
These equations describe exchange in the ensemble, however, the bias and replica exchange is also implemented in the ensemble. At a common temperature and pressure, the same acceptance rule applies when each cell is exchanged together with its configuration.
Exchange couples the replicas, so it is performed between blocks of otherwise independent dynamics. Each block advances the umbrellas for a fixed number of sequential steps using jax.lax.scan, followed by exchange attempts between neighboring umbrellas. The pairs alternate between and at successive exchange events, so no umbrella is involved in two simultaneous swaps. After the exchanges, another block of jax.lax.scan steps begins.
Appendix A.2 validates US and its REUS variant for a one-dimensional double-well system by comparing the reconstructed free energies with the exact result (Figure A1). It also gives the system, simulation setup, and postprocessing methods.

Figure A1. Double-well free energy at 300 K from standard and replica exchange umbrella sampling, each using 19 windows of 1 ns. The one-particle potential is , where is the particle position; profiles are reconstructed using the multistate Bennett acceptance ratio. Curves are means of three independent repeat profiles, using 0.0200 Å bins and a common energy zero at Å. Shading is one standard error across repeats; the dashed curve is the exact bin-integrated free energy.
Metadynamics
MTD [23] builds a time-dependent bias by adding localized positive contributions, called Gaussian hills, at the CV values visited during a simulation:
where labels a hill, is the time at which it is added, is its height, and is an unnormalized Gaussian, shifted and rescaled to vanish at a finite cutoff. Its argument is half the squared distance between the current CV vector and the hill center , measured in units of the respective hill widths. The diagonal matrix contains the squared hill widths, with entry for CV component . For periodic CVs, the corresponding components of are replaced by the shortest signed differences. When the dynamics of the system is biased by this potential, the hills raise the energy of previously visited regions over time, encouraging exploration of new regions in CV space.
Ordinary MTD adds hills of a constant height . Well-tempered MTD [24] instead reduces the height in regions where the bias has already accumulated:
where is the initial hill height, denotes the instant before the new hill is added, and is the bias factor that controls how strongly the height decreases.
The original MTD algorithm builds its bias along one trajectory and is therefore serial in time, but it can be parallelized using multiple walker MTD [27]. In this approach, several replicas, called walkers, explore the system simultaneously while contributing hills to one shared bias. The sum in Eq. (1) then includes hills from all walkers, so a region visited by one walker contributes to the bias felt by the others. With a JAX physical potential, the walkers can use the same single-GPU batching or multi-GPU sharding with local jax.vmap described for umbrellas, with a shared bias update.
Between hill additions, the walkers advance independently for a fixed number of sequential steps using jax.lax.scan. At the end of each block, every walker contributes a hill centered at its current CV vector. All (well-tempered) heights are calculated from the same bias before these hills are added, after which the shared bias and the forces are updated and the next jax.lax.scan block begins. Hill centers and heights are stored in arrays whose sizes are chosen before compilation.
Evaluating the sum directly becomes more expensive as the number of hills grows. An optional grid implementation instead stores the bias on a regular CV grid and evaluates it by interpolation between nearby grid points, following established grid representations of the bias [6]. At fixed grid resolution, this evaluation cost is independent of the number of deposited hills.
Appendix A.2 validates single walker and multiple walker well-tempered MTD against the exact free energy of the same double-well system (Figure A2). It also gives the simulation setup and postprocessing methods.

Figure A2. Double-well free energy at 300 K from single walker well-tempered metadynamics for 4 ns and four walkers sharing a bias for 1 ns each. The one-particle potential is , where is the particle position; histograms are reweighted with the metadynamics bias and its time-dependent normalization . Curves are means of three independent repeat profiles, using 0.0200 Å bins and a common energy zero at Å. Shading is one standard error across independent biases; the dashed curve is the exact bin-integrated free energy. Reweighted histograms discard the first 10% of samples while keeping all deposited hills in the bias.
Results
We now turn to three realistic systems studied in previous work, providing published free energy profiles for comparison. These examples allow us to test the implementation beyond the controlled benchmarks and demonstrate how JAX batching and execution across multiple GPUs can be used with different physical potentials and sampling methods.
Replica exchange umbrella sampling: butane torsional free energy surface
We first consider the torsional free energy profile of butane. Rotation about its central carbon–carbon bond connects the anti (trans) conformer, near a ––– torsion of , with two equivalent gauche conformers near [29]. The intervening barriers make these regions less frequently visited than the minima in unbiased dynamics, providing a simple test of sampling across the full torsional range. In Ref. 21, this profile was calculated using a spectral adaptive biasing force method [30] implemented in PySAGES with the optimized potentials for liquid simulations (OPLS) forcefield. Here we use REUS and take the curve digitized from Fig. 12 of Ref. 21 as the reference profile.
We use the JAX-MD implementation of the OPLS forcefield with parameters from the released input of Ref. 21 for flexible, 14-atom butane in vacuum, with the same temperature of 300 K and cubic cell of side 41 Å. The nonbonded cutoff is 12 Å, Lennard–Jones parameters use geometric mixing, and interactions between atoms separated by three bonds are scaled by 1 for Lennard–Jones and 1/2 for Coulomb energies. Our electrostatic settings differ from the released reference. We use Ewald electrostatics with splitting parameter 0.23 Å and reciprocal lattice indices from −10 to 10 in each direction, whereas the reference uses a mesh solver.
Our REUS calculation uses twelve periodic umbrella biases on the ––– torsion, each with force constant 5 kcal mol rad, and batches all umbrellas on one A100 GPU. Each umbrella undergoes
50 ps of equilibration and 1.00 ns of production, with exchanges attempted between neighboring umbrellas every 20 steps. Of all proposed exchanges, 41.7% are accepted.
We reconstruct the free energy profile with the weighted histogram analysis method [31], without averaging its two symmetric halves (see Figure 2). The barrier at a torsion of relative to the anti minimum is kcal mol, compared with kcal mol in the reference, while the gauche minimum lies kcal mol above anti, compared with kcal mol. After an additive energy shift, the root-mean-square (RMS) difference between the profiles is kcal mol; profiles from the first and second trajectory halves differ by kcal mol RMS. Errors are estimated from ten 100 ps blocks, leaving out one block at a time (block jackknife). These results demonstrate excellent agreement between REUS as implemented in PLUMED-JAX and the reference profile.

Figure 2. Anti (1) and gauche (2) butane conformers and the torsional free energy at 300 K using the optimized potentials for liquid simulations forcefield. Replica exchange umbrella samples are combined using the weighted histogram analysis method. The reference profile is digitized from the spectral adaptive biasing force calculation in Reference 21. Shading is one standard error. Snapshots are rendered with VMD [28].
Multiple walker metadynamics: silicon crystallization
We next consider crystallization of silicon, in which atoms rearrange from a disordered liquid into an ordered diamond crystal. In contrast to the isolated-molecule butane example, this is a bulk periodic system, and its CV depends on all atoms rather than on a four-atom torsion. Ref. [32] provides a single walker well-tempered MTD calculation using LAMMPS and PLUMED with the Stillinger-Weber potential [33]. We use its released numerical free energy profile as the reference and compare single walker and multiple walker calculations at equal total sampling time.
We simulate 216 silicon atoms using the JAX-MD implementation of the Stillinger-Weber potential at 1700 K and a target pressure of 1 bar, starting from a diamond crystal in a cubic cell of side 16.2921 Å. Following the reference, the CV is a smooth count of atoms whose neighbor arrangements resemble those in a diamond crystal, so it is small in the liquid and approaches 216 in the crystal. Each atom’s neighbors are compared with two tetrahedral reference arrangements representing the two environments in the diamond lattice [34]. The reference lattice constant is 5.431 Å, and a Gaussian width of 0.400 Å controls how much the neighbor positions can deviate from these arrangements. We combine the two similarity scores using a smooth maximum with coefficient 100, then turn the score into a continuous count using a threshold of 0.500 and a switching function with powers 6 and 12.
The timestep is 2 fs, and preparation comprises 50 ps at constant volume followed by 50 ps at constant pressure, as in the reference. Constant pressure dynamics use an isotropic barostat with a timescale of 1 ps, which expands or contracts the cell by the same factor in each direction. We thermalize the particle and volume momenta using Langevin updates with friction 1 ps; the reference instead uses stochastic velocity rescaling [35] with a 0.1 ps thermostat timescale and a 1 ps pressure timescale.
We retain the reference well-tempered MTD settings with an initial hill height of 0.622 eV, a width of 3.00 in the count of crystalline atoms, bias factor 50, and one hill per walker every 1 ps. We compare one walker run for 5.00 ns with four walkers run for 1.25 ns each, batched on an A100 GPU and contributing to a shared bias. Both calculations accumulate 5.00 ns of sampling and 5,000 hills.
Both calculations repeatedly visit the liquid and crystal states (Figure 3). We estimate the free energy from the final well-tempered bias, multiplying its negative by , where . After alignment at the liquid minimum, the RMS differences from the reference are 1.09 eV per cell for one walker and 0.656 eV per cell for four walkers. For four walkers, the liquid-to-solid barrier is 11.2 ± 0.669 eV per cell, compared with 10.8 ± 0.721 eV per cell in the reference, and the solid minimum lies 0.829 ± 1.02 eV per cell below the liquid minimum, compared with 1.36 ± 1.36 eV per cell. The shading and quoted spreads are standard deviations of eleven free energy profiles calculated at late times in each run. Overall, both calculations recover the liquid and crystal minima and a barrier that agrees with the reference within the observed late time spread.
Figure 3. Liquid (1) and diamond crystal (2) silicon configurations, and free energy profiles for the Stillinger-Weber potential at 1700 K and a target pressure of 1 bar. Profiles are estimated by scaling the final well-tempered metadynamics bias, using one or four walkers. The reference profile is the released LAMMPS/PLUMED result [32]. Shading shows the standard deviation of eleven late-time bias profiles. Snapshots are rendered with VMD [28].
Distributed umbrella sampling with a machine learning potential: lithium transport in hydrated chabazite
Finally, we consider lithium transport through hydrated chabazite, a porous crystal in which ions move between favorable sites inside water-filled channels. The free energy barriers for this motion depend on how the ion interacts with water and the surrounding crystal, as studied in Ref. [37] using the MACE–MP–0–medium machine learning potential and US. This example combines a larger solvated system with a machine learning potential, allowing us to demonstrate how umbrellas can be distributed across multiple GPUs while retaining batching within each device. We use the authors’ released umbrella data to reconstruct the reference profile [36].
We use the released 601-atom structure containing one lithium ion and 56 water molecules, retaining its fixed triclinic cell. The CV measures the lithium position along a channel. We define the channel axis using the mean positions of two eight-atom groups in the crystal framework, and project the lithium position onto this axis relative to the midpoint between the groups. Negative and positive CV values therefore describe positions on opposite sides of that midpoint.
The physical energy and forces are evaluated with the converted MACE–MP–0–medium model [38, 39] using the MACE–JAX implementation [40], while PLUMED–JAX supplies the umbrella bias. The model has two interaction layers, 128 scalar and 128 vector features, and a 6 Å cutoff; it uses cuEquivariance (cuEq) [41] for GPU acceleration. Model evaluations use 32-bit floating-point arithmetic, while integration and energy accumulation use 64-bit arithmetic.
We run 54 independent harmonic umbrellas, spaced by 0.250 Å, with force constant 89.8 kJ mol Å and a target temperature of 450 K. Each umbrella runs for 0.800 ns with a 1 fs timestep and a Langevin thermostat with friction 1 ps; the first 2 ps are discarded. The umbrella settings and duration match the released reference data. JAX sharding distributes complete umbrellas across three A100 GPUs, with jax.vmap batching 18 umbrellas on each GPU. The umbrellas evolve independently, without replica exchange.
We reconstruct both our profile and the reference using the multistate Bennett acceptance ratio [42]. After an additive energy shift, the RMS difference from the reference is 0.984 kJ mol (Figure 4). The barrier from the left minimum toward the right is 11.2 ± 0.757 kJ mol, compared with 8.71 ± in the reference, giving a difference of . The right minimum lies below the left, compared with in the reference. Errors are estimated by block averaging with eight approximately 100 ps blocks, keeping the positions of the minima and barrier fixed at those found in each full profile. Noting that these errors are biased (and thus underestimated), it may be concluded that the profiles show fair agreement in the relative stability of the two sites and the barrier between them.
Figure 4. Starting structures released by the authors near the two channel minima (1,2), with lithium in blue, and the lithium free energy profile. Umbrella sampling uses the converted MACE-MP-0-medium potential, and profiles are reconstructed using the multistate Bennett acceptance ratio. The reference profile is reconstructed from the authors’ released umbrella data [36]. Water is omitted from the images; all 56 molecules are present in the simulations. Shading is one standard error. Snapshots are rendered with VMD [28].
Computational performance
We next assess the computational cost of these calculations by measuring how much molecular dynamics (MD) can be performed in one day of wall time on one NVIDIA RTX A6000 (Figure 5). Each timing covers complete MD steps which includes building or updating neighbor lists, evaluating the physical forces and the CVs, calculating the bias forces and updating the bias when required, and advancing the positions, momenta, and temperature or pressure control. Setup, compilation, and warm-up are excluded. Butane and lithium use one umbrella, while silicon uses either one or four walkers. For several walkers, we report aggregate throughput, meaning the sum of their simulated times divided by wall time.
Figure 5. Molecular dynamics throughput on one RTX A6000, using separate linear y-axes for the three systems. Butane and lithium use one umbrella; silicon includes one and four well-tempered metadynamics walkers, with rates summed across walkers. The potentials and atom counts are shown above each panel; the reference setups follow the butane [21], silicon [32], and lithium [36] protocols described in the text. The JAX and OpenMM-ML timings for lithium both use cuEq.
For butane, the reference combines HOOMD-blue [43], which evaluates the forcefield and advances the dynamics, with PySAGES, which supplies the harmonic umbrella bias using the released butane setup of Ref. 21. For silicon, LAMMPS evaluates the Stillinger-Weber forces using its GPU implementation (Kokkos/CUDA) and advances the dynamics, while PLUMED evaluates the CV and MTD bias. For lithium, we time the authors’ released NeuralForceField (NFF) driver with MACE-MP-0-medium [36], and also OpenMM-ML, as community code that supports enhanced sampling with MLIPs, with MACE-MP-0-medium evaluated through PythonForce PyTorch MACE adapter [44] and cuEq. In the latter calculation, OpenMM evaluates the umbrella bias and advances the dynamics using CUDA kernels. The lithium JAX and OpenMM-ML calculations both use 32-bit arithmetic for MACE-MP-0-medium, 64-bit arithmetic for the bias and MD integration, and cuEq acceleration. The reference configurations are the fastest successful settings tested, using the existing software implementations.
The timesteps are 0.5 fs for butane, 2 fs for silicon, and 1 fs for lithium. HOOMD uses a 1 Å neighbor list buffer, and LAMMPS uses eight CPU threads for one silicon walker or four threads per walker for four walkers. Both silicon implementations use a bias grid initialized with 4,500 hills.
The complete JAX simulations are 3.14 times faster for butane and 5.02 times faster for single walker silicon than their measured reference configurations. Batching four silicon walkers sharing a bias increases aggregate JAX throughput from 138 to 188 ns/day, a factor of 1.36, and gives 5.06 times the throughput of the four-walker LAMMPS/PLUMED calculation. For lithium, the compiled JAX simulation reaches 14.8 ns/day, compared with 2.59 ns/day for OpenMM-ML. The JAX rate is the one-window batching control in Appendix § 8. This comparison illustrates the importance of how the force model is connected to the simulation driver: the measured OpenMM MACE adapter builds the graph on the CPU and transfers inputs and forces between CPU and GPU at every step, whereas the compiled JAX loop keeps graph construction, force evaluation, biasing, and integration on the GPU. We also attempted a like-for-like comparison using the same MACE-JAX model within OpenMM-JAX (openmmjax–cu13 version 0.1.4). The export interface does not support the cuEq operation used by this model, preventing this comparison.
To understand how several lithium umbrellas share the GPU, we compare batching with NVIDIA Multi-Process Service (MPS), which allows GPU work from separate processes to overlap (Appendix § 8). For OpenMM, this allows CPU graph construction and data transfers to overlap with GPU work from other umbrellas; for JAX, separate processes under MPS do not improve the measured batched throughput. This behavior is consistent with effective GPU use by the compiled JAX simulation: sharing the device among more umbrellas reduces the rate of each trajectory while providing a modest increase in their combined rate. Together, these comparisons demonstrate the benefit of compiling the complete simulation loop and keeping its data on the GPU.
Conclusions
PLUMED-JAX makes a small subset of PLUMED’s capabilities available to JAX-based simulation workflows. Comparisons with PLUMED establish numerical agreement for the tested CV values, derivatives, and bias calculations, while a one-dimensional double well provides an exact free energy reference for validating the sampling methods. Combining these components with JAX-MD allows force evaluation, integration, and bias updates to remain within a compiled calculation. The applications to butane and silicon demonstrate batched US with replica exchange and multiple walker MTD with a shared bias on one GPU, while lithium transport illustrates how US with a machine learning potential can be distributed across multiple GPUs. The A6000 benchmarks show the computational benefit of this approach: batching four silicon walkers increases aggregate throughput by a factor of 1.36, while lithium dynamics with one umbrella reaches 14.8 ns/day, 5.71 times the throughput of the tested OpenMM-ML configuration using the same MACE-MP-0-medium checkpoint, model precision, and cuEq acceleration. For lithium, a diagnostic comparison using JAX+MPS did not improve on batching, consistent with effective GPU use by the compiled simulation. Together, numerical validation and these execution comparisons support the use of PLUMED-JAX with JAX-MD to combine enhanced sampling, dynamics performed entirely on the GPU, and parallel simulations within a common framework.
Acknowledgments
The author thanks PLUMED’s developers and contributors for developing and maintaining the collective variables and enhanced sampling methods adapted here, and for making their implementations and documentation openly available for experimentation. The author thanks all members of the Initiative for Computational Catalysis at the Flatiron Institute for insightful discussions and Abhijeet Gangan for stimulating discussions. Computational resources were provided by the Flatiron Institute, a division of the Simons Foundation.
Data and software availability
The data supporting this work will be made available at a later stage. PLUMED-JAX is available under the Apache 2.0 license at https://github.com/venkatkapil24/plumed-jax. The JAX-MD branch used in this work is available at https://github.com/venkatkapil24/jax-md/tree/working-branch.
Numerical validation
Validation of Collective Variables
To test the collective variable (CV) implementations, we generate nine artificial point clouds containing 5, 16, or 64 atoms, using three fixed random seeds (1701, 2903, and 4309) per size. Positions are drawn uniformly from a 4 Å cube, with pair separations above 0.45 Å and noncollinear atom selections for torsions.
Each cloud is tested in four cases: without periodic boundaries, in a cubic cell of side 12 Å, in an orthorhombic cell of side lengths (12, 13, 14) Å, and in a triclinic cell with row vectors (12, 0, 0), (4.3, 11.7, 0), and (−1.1, 2.4, 13.2) Å. This gives 36 configurations. For periodic tests, the cloud is shifted by (10.1, 11.2, 12.3) Å and brought back inside the cell using the periodic boundaries.
We construct 19 further five-atom geometries by placing selected atoms at specified distances and angles to test cases unlikely to occur in random clouds. These test contact functions near a finite 0/0 limit and where they become zero (six geometries), torsions near the change from to (two), nearly straight bond angles (two), and distances where a restraining potential begins to act (three). Two geometries test distances across periodic boundaries, using a 5 Å cube or a tilted cell with row vectors (5, 0, 0), (4.6, 1.4, 0), and (0.5, 0.7, 4.8) Å. Four geometries test reference-based CVs: an exact reference, a distorted and rotated reference, and the two endpoints of a reference path.
For each applicable CV , we compare its value and the following derivatives, called gradient and deformation in Table A1:
The gradient measures the change in the CV when an atom’s position changes with the cell fixed; is component of the position of atom . The deformation derivative measures its change when the positions and cell are scaled or sheared together: contains atom positions as rows, is the identity matrix, is a dimensionless deformation matrix, and , run over , , . For nonperiodic tests, the cell argument is omitted.
The 55 configurations give 3,422 comparisons per device, covering 95 scalar outputs and option combinations; each quantity is tested only on applicable configurations. PLUMED-JAX runs on a central processing unit (CPU) and an NVIDIA RTX A6000 graphics processing unit (GPU), using JAX 0.10.2; the reference is PLUMED 2.10.1 on a CPU, with all comparisons using 64-bit floating-point arithmetic.
Table A1 gives the largest absolute differences from PLUMED over the tested configurations and, for derivatives, over their individual components. The row unit applies to value and deformation errors; gradient errors have that unit per angstrom. Summed coordination errors are reported per input atom; other quantities are not divided by atom count. The symbol 1 means dimensionless; “numeric” marks expressions without a single physical unit, evaluated with distances in angstroms. RMSD denotes root-mean-square deviation. Path coordinates and bond orientational order use the definitions of 45 and 46, respectively.
| Scalar quantity / option (unit) | CPU Value | CPU Gradient | CPU Deformation | GPU Value | GPU Gradient | GPU Deformation |
| Distance [Å] | ||||||
| Position [Å] | ||||||
| Position [Å] | ||||||
| Position [Å] | ||||||
| Center of mass–atom distance [Å] | ||||||
| Center of mass [Å] | ||||||
| Center of mass [Å] | ||||||
| Center of mass [Å] | ||||||
| Bond angle [rad] | ||||||
| Signed torsion [rad] | ||||||
| Proton transfer coordinate [Å] | ||||||
| Gyration radius, equal weights [Å] | ||||||
| Gyration radius, mass weights [Å] | ||||||
| Coordination, two groups [atom] | ||||||
| Coordination, one group [atom] | ||||||
| Coordination, overlap [atom] | ||||||
| Coordination, shifted [atom] | ||||||
| Coordination, general [atom] | ||||||
| Pair contact (1,2), raw [1] | ||||||
| Pair contact (1,2), stretched [1] | ||||||
| Pair contact (2,3), raw [1] | ||||||
| Pair contact (2,3), stretched [1] | ||||||
| Pair contact (1,4), raw [1] | ||||||
| Pair contact (1,4), stretched [1] | ||||||
| Matrix, one group, [1] | ||||||
| Matrix, one group, [1] | ||||||
| Matrix, one group, [1] | ||||||
| Matrix, one group, [1] | ||||||
| Matrix, one group, EXP [1] | ||||||
| Matrix, one group, GAUSSIAN [1] | ||||||
| Matrix, one group, TANH [1] | ||||||
| Matrix, one group, SMAP [1] | ||||||
| Matrix, one group, COSINUS [1] | ||||||
| Matrix, one group, CUBIC [1] | ||||||
| Scalar quantity / option (unit) | CPU Value | CPU Gradient | CPU Deformation | GPU Value | GPU Gradient | GPU Deformation |
| Matrix, two groups, [1] | ||||||
| Matrix, two groups, [1] | ||||||
| Matrix, two groups, [1] | ||||||
| Matrix, two groups, [1] | ||||||
| Matrix, two groups, EXP [1] | ||||||
| Matrix, two groups, GAUSSIAN [1] | ||||||
| Matrix, two groups, TANH [1] | ||||||
| Matrix, two groups, SMAP [1] | ||||||
| Matrix, two groups, COSINUS [1] | ||||||
| Matrix, two groups, CUBIC [1] | ||||||
| RMSD, translation [Å] | ||||||
| RMSD, rotation [Å] | ||||||
| Squared RMSD, translation [Ų] | ||||||
| Squared RMSD, rotation [Ų] | ||||||
| RMSD, translation, equal weights [Å] | ||||||
| RMSD, rotation, equal weights [Å] | ||||||
| Path progress [1] | ||||||
| Path distance [Ų] | ||||||
| Angular tetrahedral order [1] | ||||||
| , PLUMED normalization [1] | ||||||
| , conventional normalization [1] | ||||||
| , PLUMED normalization [1] | ||||||
| , conventional normalization [1] | ||||||
| , PLUMED normalization [1] | ||||||
| , conventional normalization [1] | ||||||
| , PLUMED normalization [1] | ||||||
| , conventional normalization [1] | ||||||
| Rational switch, removable limit [1] | ||||||
| Rational switch, raw [1] | ||||||
| Rational switch, raw cutoff [1] | ||||||
| Rational switch, stretched cutoff [1] | ||||||
| EXP switching function [1] | ||||||
| GAUSSIAN switching function [1] | ||||||
| TANH switching function [1] | ||||||
| SMAP switching function [1] | ||||||
| COSINUS switching function [1] | ||||||
| CUBIC switching function [1] |
Table A1. Maximum absolute errors for 95 scalar collective variable components, option combinations, switching functions, composition operations, and bias energies against PLUMED 2.10.1, using 64-bit floating-point arithmetic on a central processing unit and an NVIDIA RTX A6000 graphics processing unit. Horizontal rules separate collective variable types or operation families.
| Scalar quantity / option (unit) | CPU | CPU | CPU | GPU | GPU | GPU |
| Value | Gradient | Deformation | Value | Gradient | Deformation | |
| Constant scalar [1] | 0 | 0 | 0 | 0 | 0 | 0 |
| Constant vector, component 1 [1] | 0 | 0 | 0 | 0 | 0 | 0 |
| Constant vector, component 2 [1] | 0 | 0 | 0 | 0 | 0 | 0 |
| Constant vector, component 3 [1] | 0 | 0 | 0 | 0 | 0 | 0 |
| COMBINE, linear [] | ||||||
| COMBINE, mixed powers [numeric] | ||||||
| COMBINE, shifted powers [numeric] | ||||||
| COMBINE, normalized [numeric] | ||||||
| CUSTOM, arithmetic [numeric] | ||||||
| CUSTOM, [] | ||||||
| CUSTOM, sin [1] | ||||||
| CUSTOM, cos [1] | ||||||
| CUSTOM, tan [1] | ||||||
| CUSTOM, exp [1] | ||||||
| CUSTOM, log [1] | ||||||
| CUSTOM, sqrt [numeric] | ||||||
| CUSTOM, abs [] | ||||||
| CUSTOM, tanh [1] | ||||||
| CUSTOM, min [] | ||||||
| CUSTOM, max [] | ||||||
| Harmonic distance bias [kJ mol] | ||||||
| Harmonic torsion bias [kJ mol] | ||||||
| Upper wall bias [kJ mol] | ||||||
| Lower wall bias [kJ mol] |
Table A1 (continued).
Validation of umbrella sampling and metadynamics
We test umbrella sampling (US) and metadynamics (MTD) for one particle of mass 12 atomic mass units in the one-dimensional potential , where is the particle’s Cartesian position along the sampled direction. The exact free energy is , where sets the arbitrary energy zero. It has equal minima at Å and a barrier of . All four sampling variants use JAX-MD to integrate Langevin dynamics at 300 K with a 1 fs timestep and friction [19]. Standard and replica exchange US (REUS) each use 19 windows spaced by 0.200 Å from to Å, with stiffness , 20 ps of equilibration, and 1 ns of production per window. The exchange variant attempts swaps between alternating neighboring pairs every 20 fs and refreshes the forces after accepted swaps; the acceptance fraction is 0.589. Three independently seeded window sets are batched together, and the free energy profile for each set is reconstructed separately using the multistate Bennett acceptance ratio.
We compare single walker well-tempered MTD for 4 ns with four walkers sharing one bias for 1 ns each, giving the same aggregate sampling time of 4 ns per independent repeat. Three independently seeded repeats use 20 ps of unbiased equilibration, hill width 0.200 Å, initial height per hill, bias factor 5, and deposition every 20 fs per walker. Thus, the per-hill parameters and nominal hill counts match at equal aggregate sampling, while the actual well-tempered heights depend on the evolving bias. The bias grid spans Å with 1,200 intervals. MTD profiles use the recorded bias and weights [47] after discarding the initial 10% of sampled configurations; all deposited hills remain in the bias.
Figures A1 and A2 compare mean profiles from the three independent repeats, using 0.0200 Å bins and a common energy zero at Å. The root-mean-square errors relative to the exact bin-integrated profile, over coordinates with , are 0.0627 and 0.0813 for standard and replica exchange umbrellas, and 0.0658 and 0.0702 for single walker and multiple walker MTD, respectively.
Parallel execution of lithium umbrellas
Figure A3 compares batched JAX-MD using jax.vmap with separate OpenMM-ML or JAX-MD processes under NVIDIA Multi-Process Service (MPS), for one, two, or four independent lithium umbrellas on one RTX A6000. The single-window JAX control also uses jax.vmap. The JAX+MPS calculation is used as a diagnostic of GPU usage; its driver is not included in the repository. No calculation in this comparison uses replica exchange.
Figure A3. Throughput per lithium umbrella on one RTX A6000, obtained by dividing aggregate throughput by the number of windows. All cases use the MACE-MP-0-medium potential for 601 atoms and independent harmonic umbrellas at 450 K, without replica exchange.
With batched JAX, the rate per umbrella decreases from 14.8 to 4.42 ns/day as the number of windows increases from one to four; aggregate throughput increases by a factor of 1.19. For OpenMM+MPS, the corresponding rates are 2.71 and 1.85 ns/day. JAX+MPS gives 3.92 ns/day per umbrella at four windows, below the batched JAX result.
All cases use the same MACE-MP-0-medium checkpoint, 32-bit model inference, 64-bit bias and integration, a 1 fs timestep, a 450 K Langevin thermostat with friction , and umbrella stiffness , with cuEquivariance acceleration. Window centers are spaced by , symmetrically around the CV value in the common starting structure. Each process has four CPU cores available, with Torch and numerical library thread counts set to one.
Each window is timed over 100 steps after warm-up, excluding trajectory output. Aggregate throughput is the sum of simulated time across windows divided by the elapsed time from the first worker’s start to the last worker’s finish.
References
- [1]Mark Tuckerman. Statistical Mechanics: Theory and Molecular Simulation. OUP Oxford, February 2010. ISBN 9780191523465.
- [2]Massimo Bocus, Elias Van Den Broeck, Xian Wu, Mathias Bal, Jeroen Bomon, Louis Vanduijfhuys, Bert F. Sels, Bert U. W. Maes, and Veronique Van Speybroeck. Confined hot-pressurized water in Brønsted-acidic beta zeolite speeds up the O-demethylation of guaiacol. Nature Catalysis, 8(1):33–45, January 2025. ISSN 2520-1158. doi: 10.1038/s41929-024-01282-6. URL https://www.nature.com/articles/s41929-024-01282-6.DOI
- [3]Zoe L. Watson, Isaac J. Knudson, Fred R. Ward, Scott J. Miller, Jamie H. D. Cate, Alanna Schepartz, and Ara M. Abramyan. Atomistic simulations of the Escherichia coli ribosome provide selection criteria for translationally active substrates. Nature Chemistry, 15(7):913–921, July 2023. ISSN 1755-4349. doi: 10.1038/s41557-023-01226-w. URL https://www.nature.com/articles/s41557-023-01226-w.DOI
- [4]Mingfeng Liu, Jiantao Wang, Junwei Hu, Peitao Liu, Haiyang Niu, Xuexi Yan, Jiangxu Li, Haile Yan, Bo Yang, Yan Sun, Chunlin Chen, Georg Kresse, Liang Zuo, and Xing-Qiu Chen. Layer-by-layer phase transformation in Ti3O5 revealed by machine-learning molecular dynamics simulations. Nature Communications, 15(1):3079, April 2024. ISSN 2041-1723. doi: 10.1038/s41467-024-47422-1. URL https://www.nature.com/articles/s41467-024-47422-1.DOI
- [5]Venkat Kapil, Christoph Schran, Andrea Zen, Ji Chen, Chris J. Pickard, and Angelos Michaelides. The first-principles phase diagram of monolayer nanoconfined water. Nature, 609(7927):512–516, September 2022. ISSN 1476-4687. doi: 10.1038/s41586-022-05036-x. URL https://www.nature.com/articles/s41586-022-05036-x.DOI
- [6]Massimiliano Bonomi, Davide Branduardi, Giovanni Bussi, Carlo Camilloni, Davide Provasi, Paolo Raiteri, Davide Donadio, Fabrizio Marinelli, Fabio Pietrucci, Riccardo A. Broglia, and Michele Parrinello. PLUMED: A portable plugin for free-energy calculations with molecular dynamics. Computer Physics Communications, 180(10):1961–1972, October 2009. ISSN 0010-4655. doi: 10.1016/j.cpc.2009.05.011. URL https://www.sciencedirect.com/science/article/pii/S001046550900157X.DOI
- [7]Gareth A. Tribello, Massimiliano Bonomi, Davide Brunduardi, Carlo Camilloni, and Giovanni Bussi. PLUMED 2: New feathers for an old bird. Computer Physics Communications, 185(2):604–613, February 2014. ISSN 0010-4655. doi: 10.1016/j.cpc.2013.09.018. URL https://www.sciencedirect.com/science/article/pii/S0010465513003196.DOI
- [8]The PLUMED consortium. Promoting transparency and reproducibility in enhanced molecular simulations. Nature Methods, 16(8):670–673, August 2019. ISSN 1548-7105. doi: 10.1038/s41592-019-0506-8. URL https://www.nature.com/articles/s41592-019-0506-8.DOI
- [9]Daniele Rapetti, Massimiliano Bonomi, Carlo Camilloni, Giovanni Bussi, and Gareth A. Tribello. Making PLUMED fly: A tutorial on optimizing performance. The Journal of Physical Chemistry B, 130(9):2504–2517, March 2026. ISSN 1520-6106. doi: 10.1021/acs.jpcb.5c07562. URL https://doi.org/10.1021/acs.jpcb.5c07562.
- [10]Jim Pfaendtner and Massimiliano Bonomi. Efficient sampling of high-dimensional free-energy landscapes with parallel bias metadynamics. Journal of Chemical Theory and Computation, 11(11):5062–5067, November 2015. ISSN 1549-9618. doi: 10.1021/acs.jctc.5b00846. URL https://doi.org/10.1021/acs.jctc.5b00846.
- [11]Omar Valsson and Michele Parrinello. Variational approach to enhanced sampling and free energy calculations. Physical Review Letters, 113(9):090601, August 2014. doi: 10.1103/PhysRevLett.113.090601. URL https://link.aps.org/doi/10.1103/PhysRevLett.113.090601.DOI
- [12]Michele Invernizzi and Michele Parrinello. Rethinking metadynamics: From bias potentials to probability distributions. The Journal of Physical Chemistry Letters, 11(7):2731–2736, April 2020. ISSN 1948-7185. doi: 10.1021/acs.jpclett.0c00497. URL https://doi.org/10.1021/acs.jpclett.0c00497.
- [13]Szilárd Páll, Artem Zhmurov, Paul Bauer, Mark Abraham, Magnus Lundborg, Alan Gray, Berk Hess, and Erik Lindahl. Heterogeneous parallelization and acceleration of molecular dynamics simulations in GROMACS. The Journal of Chemical Physics, 153(13):134110, October 2020. ISSN 0021-9606. doi: 10.1063/5.0018516. URL https://doi.org/10.1063/5.0018516.
- [14]Aidan P. Thompson, H. Metin Aktulga, Richard Berger, Dan S. Bolintineanu, W. Michael Brown, Paul S. Crozier, Pieter J. in ’t Veld, Axel Kohlmeyer, Stan G. Moore, Trung Dac Nguyen, Ray Shan, Mark J. Stevens, Julien Tranchida, Christian Trott, and Steven J. Plimpton. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer Physics Communications, 271:108171, February 2022. ISSN 0010-4655. doi: 10.1016/j.cpc.2021.108171. URL https://www.sciencedirect.com/science/article/pii/S0010465521002836.DOI
- [15]Peter Eastman, Raimondas Galvelis, Raúl P. Peláez, Charles R. A. Abreu, Stephen E. Farr, Emilio Gallicchio, Anton Gorenko, Michael M. Henry, Frank Hu, Jing Huang, Andreas Krämer, Julien Michel, Joshua A. Mitchell, Vijay S. Pande, João PGLM Rodrigues, Jaime Rodriguez-Guerra, Andrew C. Simmonett, Sukrit Singh, Jason Swails, Philip Turner, Yuanqing Wang, Ivy Zhang, John D. Chodera, Gianni De Fabritiis, and Thomas E. Markland. OpenMM 8: Molecular dynamics simulation with machine learning potentials. The Journal of Physical Chemistry B, 128(1):109–116, January 2024. ISSN 1520-6106. doi: 10.1021/acs.jpcb.3c06662. URL https://doi.org/10.1021/acs.jpcb.3c06662.
- [16]Stefan Seritan, Christoph Bannwarth, Bryan S. Fales, Edward G. Hohenstein, Christine M. Isborn, Sara I. L. Kokkila-Schumacher, Xin Li, Fang Liu, Nathan Luehr, James W. Snyder Jr., Chenchen Song, Alexey V. Titov, Ivan S. Ufimtsev, Lee-Ping Wang, and Todd J. Martínez. TeraChem: A graphical processing unit-accelerated electronic structure package for large-scale ab initio molecular dynamics. WIREs Computational Molecular Science, 11(2):e1494, 2021. ISSN 1759-0884. doi: 10.1002/wcms.1494. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/wcms.1494.DOI
- [17]Rui Li, Qiming Sun, Xing Zhang, and Garnet Kin-Lic Chan. Introducing GPU acceleration into the Python-based simulations of chemistry framework. The Journal of Physical Chemistry A, 129(5):1459–1468, February 2025. ISSN 1089-5639. doi: 10.1021/acs.jpca.4c05876. URL https://doi.org/10.1021/acs.jpca.4c05876.
- [18]Marcella Iannuzzi, Jan Wilhelm, Frederick Stein, Augustin Bussy, Hossam Elgabarty, Dorothea Golze, Anna-Sophia Hehn, Maximilian Graml, Stepan Marek, Beliz Sertcan Gökmen, Christoph Schran, Harald Forbert, Rustam Z. Khaliullin, Anton Kozhevnikov, Mathieu Taillefumier, Rocco Meli, Vladimir V. Rybkin, Martin Brehm, Robert Schade, Ole Schütt, Johann V. Pototschnig, Hossein Mirhosseini, Andreas Knüpfer, Dominik Marx, Matthias Krack, Jürg Hutter, and Thomas D. Kühne. The CP2K program package made simple. The Journal of Physical Chemistry B, 130(4):1237–1310, January 2026. ISSN 1520-6106. doi: 10.1021/acs.jpcb.5c05851. URL https://doi.org/10.1021/acs.jpcb.5c05851.
- [19]Samuel S. Schoenholz and Ekin D. Cubuk. JAX MD: A framework for differentiable physics. In Advances in Neural Information Processing Systems, volume 33, pages 11428–11441. Curran Associates, Inc., 2020. URL https://proceedings.neurips.cc/paper/2020/hash/83d3d4b6c9579515e1679acac8cbc8033-Abstract.html.arxiv.org/abs/1912.04232
- [20]Orion Cohen, Janosh Riebesell, Rhys Goodall, Adeesh Kolluru, Stefano Falletta, Joseph Krause, Jorge Colindres, Gerbrand Ceder, and Abhijeet S. Gangan. TorchSim: An efficient atomistic simulation engine in PyTorch. AI for Science, 1(2):0250003, 2025. doi: 10.1088/3050-287x/ae1799. URL https://doi.org/10.1088/3050-287x/ae1799.
- [21]Pablo F. Zubieta Rico, Ludwig Schneider, Gustavo R. Pérez-Lemus, Riccardo Alessandri, Siva Dasetty, Trung D. Nguyen, Cintia A. Menéndez, Yiheng Wu, Yezhi Jin, Yinan Xu, Samuel Varner, John A. Parker, Andrew L. Ferguson, Jonathan K. Whitmer, and Juan J. de Pablo. PySAGES: flexible, advanced sampling methods accelerated with GPUs. npj Computational Materials, 10(1):35, February 2024. ISSN 2057-3960. doi: 10.1038/s41524-023-01189-z. URL https://www.nature.com/articles/s41524-023-01189-z.
- [22]G. M. Torrie and J. P. Valleau. Nonphysical sampling distributions in Monte Carlo free-energy estimation: Umbrella sampling. Journal of Computational Physics, 23(2):187–199, February 1977. ISSN 0021-9991. doi: 10.1016/0021-9991(77)90121-8. URL https://www.sciencedirect.com/science/article/pii/0021999177901218.DOI
- [23]Alessandro Laio and Michele Parrinello. Escaping free-energy minima. Proceedings of the National Academy of Sciences, 99(20):12562–12566, 2002. doi: 10.1073/pnas.202427399.DOI
- [24]Alessandro Barducci, Giovanni Bussi, and Michele Parrinello. Well-tempered metadynamics: A smoothly converging and tunable free-energy method. Physical Review Letters, 100(2):020603, January 2008. doi: 10.1103/PhysRevLett.100.020603. URL https://link.aps.org/doi/10.1103/PhysRevLett.100.020603.DOI
- [25]James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Yash Katariya, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs. Software repository, 2018. URL https://github.com/jax-ml/jax. Accessed 30 October 2026.
- [26]Yuji Sugita, Akio Kitao, and Yuko Okamoto. Multidimensional replica-exchange method for free-energy calculations. The Journal of Chemical Physics, 113(15):6042–6051, 2000. doi: 10.1063/1.1308516.DOI
- [27]Paolo Raiteri, Alessandro Laio, Francesco Luigi Gervasio, Cristian Micheletti, and Michele Parrinello. Efficient reconstruction of complex free energy landscapes by multiple walkers metadynamics. The Journal of Physical Chemistry B, 110(8):3533–3539, 2006. doi: 10.1021/jp054359r. URL https://doi.org/10.1021/jp054359r.
- [28]William Humphrey, Andrew Dalke, and Klaus Schulten. VMD: Visual molecular dynamics. Journal of Molecular Graphics, 14(1):33–38, 1996. doi: 10.1016/0263-7855(96)00018-5.DOI
- [29]William L. Jorgensen, David S. Maxwell, and Julian Tirado-Rives. Development and testing of the OPLS all-atom force field on conformational energetics and properties of organic liquids. Journal of the American Chemical Society, 118(45):11225–11236, 1996. doi: 10.1021/ja962176o.DOI
- [30]Pablo F. Zubieta Rico, Gustavo R. Pérez-Lemus, and Juan J. de Pablo. Efficient sampling of free energy landscapes with functions in Sobolev spaces. The Journal of Chemical Physics, 162(8):084109, 2025. doi: 10.1063/5.0221263. URL https://doi.org/10.1063/5.0221263.
- [31]Shankar Kumar, Djamel Bouzida, Robert H. Swendsen, Peter A. Kollman, and John M. Rosenberg. The weighted histogram analysis method for free-energy calculations on biomolecules. I. the method. Journal of Computational Chemistry, 13(8):1011–1021, 1992. doi: 10.1002/jcc.540130812.
- [32]Pablo M. Piaggi. Tutorial: Crystallization of silicon. Input files and numerical results for LAMMPS/PLUMED simulations, n.d. URL https://github.com/PabloMPiaggi/Crystallization-of-Silicon. Metadynamics-1700K example; accessed 30 October 2026.
- [33]Frank H. Stillinger and Thomas A. Weber. Computer simulation of local order in condensed phases of silicon. Physical Review B, 31(8):5262–5271, 1985. doi: 10.1103/PhysRevB.31.5262.DOI
- [34]Pablo M. Piaggi and Michele Parrinello. Calculation of phase diagrams in the multithermal-multibaric ensemble. The Journal of Chemical Physics, 150(24):244119, 2019. doi: 10.1063/1.5102104.DOI
- [35]Giovanni Bussi, Tatyana Zykova-Timan, and Michele Parrinello. Isothermal-isobaric molecular dynamics using stochastic velocity rescaling. The Journal of Chemical Physics, 130(7):074101, 2009. doi: 10.1063/1.3073889.DOI
- [36]Sauradeep Majumdar, Swagata Roy, Kyujung Jun, Miguel Steiner, and Rafael Gómez-Bombarelli. Dataset: The free energy landscape of Li- and Na-ion transport in nanoconfinement with machine learning interatomic potentials. Dataset, Zenodo, 2026. URL https://zenodo.org/records/18511449.
- [37]Sauradeep Majumdar, Swagata Roy, Kyujung Jun, Miguel Steiner, and Rafael Gómez-Bombarelli. The free energy landscape of Li- and Na-ion transport in nanoconfinement with machine learning interatomic potentials. Chemistry of Materials, 38(8):3971–3981, 2026. doi: 10.1021/acs.chemmater.5c02940.
- [38]Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Systems, volume 35, pages 11423–11436. Curran Associates, Inc., 2022. doi: 10.52202/068431-0830. URL https://papers.neurips.cc/paper_files/paper/2022/hash/4a36c3c51aff11ed9f34615b81edb5bbc-Abstract-Conference.html.DOI
- [39]Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, Dávid P. Kovács, Janosh Riebesell, Xavier R. Advincula, Mark Asta, Matthew Aylott, William J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Filippo Bigi, Samuel M. Blau, Vlad Cǎrare, Michele Ceriotti, Sangyu Chong, James P. Darby, Sandip De, Flaviano Della Pia, Volker L. Deringer, Rokas Elišiojus, Zakariya El-Machachi, Edvin Fako, Fabio Falconi, Andrea C. Ferrari, John L. A. Gardner, Mikolaj J. Gawkowski, Annelena Genreith-Schriever, Janine George, Rhys E. A. Goodall, Jonas Grandel, Clare P. Grey, Petr Grigorev, Shuang Han, Will Handley, Hendrik H. Heenen, Kersti Hermansson, Cheuk Hin Ho, Stephan Hofmann, Christian Holm, Jad Jaafar, Konstantin S. Jakob, Hyunwook Jung, Venkat Kapil, Aaron D. Kaplan, Nima Karimitar, James R. Kermode, Panagiotis Kourits, Namu Kroupa, Jolla Kullgren, Matthew C. Kuner, Kuryla, Guoda Liepouniute, Chen Lin, Johannes T. Margraf, Ioan-Bogdan Magdău, Angelos Michaelides, J. Harry Moore, Akash A. Naik, Samuel P. Niblett, Sam Walton Norwood, Niamh O’Neill, Christoph Ortner, Kristin A. Persson, Karsten Reuter, Andrew S. Rosen, Louise A. M. Rosset, Lars L. Schaaf, Christoph Schran, Benjamin X. Shi, Eric Sivonxay, Tamás K. Stenczel, Christopher Sutton, Viktor Svahn, Thomas D. Swinburne, Jules Tilly, Cas van der Oord, Santiago Vargas, Eszter Varga-Umbrich, Tejs Vegge, Martin Vondrák, Yangshuai Wang, William C. Witt, Thomas Wolf, Fabian Zills, and Gábor Csányi. A foundation model for atomistic materials chemistry. The Journal of Chemical Physics, 163(18):184110, 2025. doi: 10.1063/5.0297006.DOI
- [40]The MACE-JAX developers. MACE-JAX: Equivariant machine learning interatomic potentials in JAX. Software repository, n.d. URL https://github.com/ACESuit/mace-jax. Accessed 30 October 2026.
- [41]NVIDIA. cuEquivariance. Software repository, n.d. URL https://github.com/NVIDIA/cuEquivariance. Accessed 30 October 2026.
- [42]Michael R. Shirts and John D. Chodera. Statistically optimal analysis of samples from multiple equilibrium states. The Journal of Chemical Physics, 129(12):124105, 2008. doi: 10.1063/1.2978177.DOI
- [43]Joshua A. Anderson, Jens Glaser, and Sharon C. Glotzer. HOOMD-blue: A Python package for high-performance molecular dynamics and hard particle Monte Carlo simulations. Computational Materials Science, 173:109363, 2020. doi: 10.1016/j.commatsci.2019.109363.DOI
- [44]Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, volume 32, pages 8024–8035. Curran Associates, Inc., 2019. URL https://proceedings.neurips.cc/paper/2019/hash/bd bca288fee179212bfa9f7012727740-Abstract.html.arxiv.org/abs/1912.01703
- [47]Pratyush Tiwary and Michele Parrinello. A time-independent free energy estimator for metadynamics. The Journal of Physical Chemistry B, 119(3):736–742, 2015. doi: 10.1021/jp504920s.DOI




