AceFF: A State-of-the-Art Machine Learning Potential for Small Molecules
Authors: Stephen E. Farr, Stefan Doerr, Antonio Mirarchi, Alireza Moradzadeh, Francesc Sabanes Zariquiey, Saee Paliwal, Gianni De Fabritiis
Organizations: Acellera Labs, C Dr Trueta 183, 08005, Barcelona, Spain · Computational Science Laboratory, Universitat Pompeu Fabra, Barcelona Biomedical Research Park (PRBB), C Dr. Aiguader 88, 08003, Barcelona, Spain · Institució Catalana de Recerca i Estudis Avançats (ICREA), Passeig Lluis Companys 23, 08010 Barcelona, Spain
We introduce AceFF, a pre-trained machine learning interatomic potential (MLIP) optimized for small molecule drug discovery. While MLIPs have emerged as efficient alternatives to Density Functional Theory (DFT), generalizability across diverse chemical spaces remains difficult. AceFF addresses this by refining the TensorNet2 architecture and training it on a comprehensive dataset of drug-like compounds. This approach yields a force field that balances high-throughput inference speed with DFT-level accuracy. We introduce two models sharing the same architecture: \mbox{AceFF-2}, trained on our own dataset of small molecules and supporting the essential medicinal chemistry elements (H, B, C, N, O, F, Si, P, S, Cl, Br, I); and \mbox{AceFF-2-OMOL}, trained on a subset of OMol25 containing the AceFF-2 elements plus the common biomolecular ion elements Na and K. The models are explicitly trained to handle charged states. Validation against rigorous benchmarks, including complex torsional energy scans, molecular dynamics trajectories, batched minimizations, and tests of force and energy accuracy, demonstrates that AceFF is state-of-the-art for organic molecules in the accuracy and speed regime important for drug discovery. The AceFF-2 and AceFF-2-OMol model weights and inference code are available at https://huggingface.co/Acellera/AceFF-2.0 and https://huggingface.co/Acellera/AceFF-2-OMOL.
Figures & tables
Figure 1: The AceFF-2 architecture TensorNet2 builds upon TensorNet Simeon and De Fabritiis (2023) with the new components colored in blue. The main change is that after the tensor embedding and each tensor interaction block, a set of partial charges q and weights w are computed and undergo a neutral charge equilibration (NQE) procedure as done in AIMNet2 Anstine et al. (2025) . The partial charges are then combined with edge features during the tensor interaction. The predicted charges are used to compute a Coulomb energy ECoulomb alongside the standard short-range node-wise MLIP energy term Ei . The dashed arrow represents where iteration occurs for more than one interaction layer. The pictured diagram corresponds to TensorNet2 1-layer. The grey dashed circles around the example molecule illustrate the short cutoffs used in the tensor embedding and interaction. The blue dashed circle represents the long-range of the total charge (a global feature), the NQE block, and the Coulomb interaction.
Figure 2: Molecular dynamics speed (time per step) of AceFF-2 for increasing number of atoms, using periodic water boxes with a reaction-field Coulomb cutoff of 1.2 nm (or half the box length for boxes smaller than 2.4 nm), on an RTX 4090. The blue line is torchmd-net with the pure PyTorch TensorNet ops, the orange line is torchmd-net version 3 with the Warp kernels, and the green line is the acellera-tensornet version.
Figure 3: Torsion scans. a. Sellers et al. torsion scan benchmark Sellers et al. (2017) . The orange lines are the median value. The methods are ordered from left to right by median MAE compared to the coupled cluster reference data. The y-axis has been intentionally capped at 2.0 kcal/mol to focus on the differences between the MLIP models. An image of one of the molecules with the torsion angle indicated is pictured. Values marked by * were taken from a previous publication Sabanés Zariquiey et al. (2025) . All other data points were recalculated. b. Behara et al torsion scan benchmark Behara et al. (2024) . The orange lines are the median value. The methods are ordered from left to right by median MAE compared to the coupled cluster reference data. The scatter points are color-coded by the molecular charge. The y-axis has been intentionally capped at 4.0kcal/mol to focus on the differences between the MLIPs. Data points of methods marked with * were taken from the original data Behara et al. (2024) , all other data points were recalculated. On both plots, a guiding line has been drawn at 1 kcal/mol.
Method
MAE (kcal/mol)
RMSE (kcal/mol)
GAFF
22.87
30.86
GFN2-XTB
14.6
15.2
ANI-2X
4.41
5.41
g-XTB
3.85
4.76
AceFF-1.0
2.73
3.32
AIMNet2
2.39
3.13
Table 1: Performance of MLIPs, two semi-empirical methods, and one MM force field on the Wiggle150 benchmark Brew et al. (2025) .
Figure 4: Schrödinger ligand test set evaluation. a. Energy error vs force error for AceFF-1.0, color-coded by ligand total charge. b. Energy error vs force error for AceFF-2, color-coded by ligand total charge. The axis range is the same in sub-plots a and b to aid comparison. The inset graph in sub-plot b shows a zoomed-in view of the data points. c. Force errors for all tested models. The orange line is the median value. The data points are color-coded by charge.
Figure 5: SPICE test set performance. a. Overall MAE plotted against the number of model parameters as in Eastman et al. (2026) . All gray data points are taken from Eastman et al. (2026) . Points are categorized into models trained on OMol25 (stars) and models trained on pre-OMol25 datasets (circles). b. Overall MAE plotted against simulation speed (steps per second for a 2,661-atom water box) as done in Eastman et al. (2026) , recomputed on an RTX 4090 GPU. The blue star denotes AceFF-2-OMol using the TorchMD-NET implementation (with Warp kernels), and the green star denotes the Acellera-TensorNet plugin. Points marked with a † are MACE models evaluated with randomly initialized weights using architectures corresponding to the released model variants.
Figure 6: PLA15 benchmark evaluation. Box plots show the distribution of signed percentage errors across all evaluated models.
Figure 7: Potential energy of ethane C-C bond length scan for different MLIPs compared to a reference of DFT with ω B97m-V/tzvppd.
Figure 8: Inference speed in steps per second for different models versus the number of atoms. We use the water box speed benchmarks from Eastman et al. (2026) .
Figure 9: Batched molecular dynamics with AceFF-2. a. The total time per step for the corresponding number of conformers in batched mode (blue line) and in serial with OpenMM (black line). b. The effective total throughput in ns/day for batched mode (red line) and using OpenMM in serial (black line).
Figure 10: Ligand RMSD of NVT molecular dynamics of protein-ligand complex using ML/MM. a. Visualization system; the MLIP region has been highlighted, and water has been omitted for clarity. b. The RMSD of the ligand is plotted for the 1fs and 2fs timestep AceFF-2 simulations in blue and orange, respectively. c. The ligand-pocket center of mass distance is plotted for the 1fs and 2fs timestep AceFF-2 simulations in blue and orange, respectively.
MLIP
mean
std
median
% < 1 Å
% < 2 Å
% < 3 Å
% < 10 Å
Number failed
AceFF-2
0.551
0.447
0.419
86.29
98.43
99.72
99.93
2
AceFF-2-OMol
0.527
0.446
0.392
87.20
98.46
99.58
99.76
7
AIMNet2
0.508
0.426
0.388
89.75
98.60
99.62
99.72
8
OrbMol
0.534
0.433
0.400
86.92
98.64
99.86
100.00
0
Table 2: Platinum Diverse geometry optimization benchmark Friedrich et al. (2017) . RMSD (Å) between the reference crystallographic ligand conformation and the MLIP-minimized structure. Mean, standard deviation, median, and cumulative percentages are computed over the conformers for which all four methods succeeded. “Number failed” is the number of conformers with an RMSD above 10 Å (exploded structures).
Figure 11: Platinum Diverse geometry optimization benchmark. a. The histogram of the RMSDs between the initial reference structure and the geometry-optimized structure for the four tested MLIPs for all conformers in the Platinum Diverse set Friedrich et al. (2017) . b. Molecule index 1653 has a net charge of −8 due to four phosphate groups. When minimized with AceFF-2 the phosphates are ejected. c. Molecule index 1816 has a net charge of −3 and two phosphate groups; when minimized with AceFF-2 one of the phosphates is ejected. d. Molecule index 1928 has the largest RMSD (3.9 Å) for AceFF-2; it has a net charge of −4 , and upon minimization with AceFF-2 there is a change in bond topology as highlighted by the red circles.
Figure 12: Comparison of batched conformer minimization speed with torch.optim.LBFGS and ASE serial minimization of the same number of conformers of an Imatinib molecule.
We introduce implicit machine learning force fields (I-MLFFs), which replace explicit stacks of neural network layers with self-consistent fixed-point equations. In molecular simulations, this formulation enables intermediate representations to be reused across successive timesteps, thereby warm-starting force evaluation. The resulting models effectively combine the computational footprint of a shallow, single-layer MLFF with the representational capacity and accuracy of a deep neural network. Our approach unlocks architecture-agnostic efficiency gains that are inaccessible when force prediction and trajectory integration are considered separately. We demonstrate this across three major classes of graph neural networks: invariant, equivariant Cartesian tensor, and SO(3)-equivariant spherical-tensor architectures. Each yields a two- to five-fold reduction in compute and memory footprint. Crucially, these gains are achieved while retaining full atomistic resolution and the original integration timestep, avoiding spatial or temporal coarse graining. Our contribution therefore advances the scaling frontier of quantum-mechanically faithful molecular simulation, enabling longer trajectories and larger atomistic systems within fixed GPU memory and compute budgets, and thereby opening access to new insights across biomolecular and material systems.
Johannes Maeß, Leon Werner, J. Thorben Frank +5
BIFOLD, Berlin, Germany. · Machine Learning Group, TU Berlin, Berlin, Germany. · Department of Artificial Intelligence, Korea University, Seoul, South Korea. +1
General-purpose large language models (LLMs) that rely on in-context learning do not reliably deliver the scientific understanding and performance required for drug discovery tasks. Simply increasing model size or introducing reasoning tokens does not yield significant performance gains. To address this gap, we introduce the MMAI Gym for Science, a one-stop shop molecular data formats and modalities as well as task-specific reasoning, training, and benchmarking recipes designed to teach foundation models the 'language of molecules' in order to solve practical drug discovery problems. We use MMAI Gym to train an efficient Liquid Foundation Model (LFM) for these applications, demonstrating that smaller, purpose-trained foundation models can outperform substantially larger general-purpose or specialist models on molecular benchmarks. Across essential drug discovery tasks - including molecular optimization, ADMET property prediction, retrosynthesis, drug-target activity prediction, and functional group reasoning - the resulting model achieves near specialist-level performance and, in the majority of settings, surpasses larger models, while remaining more efficient and broadly applicable in the domain.
Maksim Kuznetsov, Zulfat Miftahutdinov, Rim Shayakhmetov +17
Machine learning interatomic potentials (MLIPs) achieve excellent accuracy when trained on large Density Functional Theory (DFT) data. To be useful in practice, they must often be adapted to target chemistries using small and expensive task-specific datasets. However, MLIPs transfer inconsistently across domains, with representations that often loose accessible composition and structure information. To address this, we present TriForces, a model-agnostic three-stream framework that separates composition and structure information, combined with self-supervised learning to preserve transferable representations. TriForces improves performance on MatBench and QM9 over baselines without needing DFT labels and enables efficient similar structure retrieval through its learned latent space. On OMat24, in limited-data training regime, TriForces reduces energy MAE by 57% at 20K samples only and improves force MAE across sample sizes. We release pretrained TriForces variants across multiple MLIP architectures with code at https://github.com/Ramlaoui/triforces.
Ali Ramlaoui, Alexandre Duval, Hannah Bull +4
Entalpic, Paris, France · Universit´e Paris-Saclay, CentraleSup´elec, Inria, Gif-sur-Yvette, France