DensIP is introduced, a physics-based model of intermolecular interactions that uses machine-learned electron densities and only four universal parameters that outperforms state-of-the-art general-purpose MLFFs for long-range interactions and can be applied to molecules as large as drug ligands.
Abstract
Machine-learned force fields (MLFFs) contain many learnable parameters and therefore require large training datasets. This poses a challenge for developing highly accurate, general-purpose MLFFs because generating high-quality ab initio reference data is computationally expensive. Classical empirical potentials offer a potentially inexpensive source of synthetic training data, but existing models often lack the accuracy needed to provide useful reference energies. Here, we introduce the density-based intermolecular potential (DensIP), a physics-based model of intermolecular interactions that uses machine-learned electron densities and only four universal parameters. We train and test DensIP on CCSD(T)/CBS interaction energies from DES15K, a dataset of dimers of small organic molecules. DensIP achieves sub-kcal/mol errors for dimers containing molecules absent from the training set, including molecules in non-equilibrium conformations, demonstrating strong transferability. We further show that DensIP can be applied to molecules as large as drug ligands. Notably, DensIP outperforms state-of-the-art general-purpose MLFFs for long-range interactions, making it a promising approach for generating accurate synthetic training data at scale.
Machine learning interatomic potentials (MLPs) have revolutionized atomistic modeling, offering the potential to replace traditional methods like Density Functional Theory (DFT). However, inference time of MLPs is orders of magnitude slower than that of classical force fields, hindering real-world applications for biomolecular systems that require timescales of microseconds and beyond. Implicit solvent MLPs can address this issue, but are faced with data challenges associated with coarse-grained modeling. Consequently, previous approaches relied on empirical force field data, thereby inherently limiting the MLP's accuracy. Here, we introduce the Transferable Water Implicit Network (TWIN), an implicit water MLP parametrized entirely by an Equivariant Graph Neural Network and trained solely on ab initio and experimental labels. We demonstrate TWIN's transferability across drug-like molecules, peptides, and proteins, achieving excellent results on ab initio and experimental crystallographic and NMR benchmarks, consistently outperforming previous machine-learning-based implicit solvent or coarse-grained models. Furthermore, TWIN closely matches DFT-based explicit solvent MLPs while providing a two-order-of-magnitude faster timestep evaluation, paving the way for efficient ab initio-level modeling of biomolecular systems in aqueous environments.
This work demonstrates how recent foundational machine learning interatomic potentials (MLIPs) trained at the r$^2$SCAN level can be leveraged to improve the agreement of formation energies with experiment, reducing the mean absolute error by more than 40% relative to GGA without requiring any additional DFT calculation.
Timo Reents, Marnik Bercx, Giovanni Pizzi· 0 citations
Pretrained machine-learning interatomic potentials, so-called universal or foundation models offer an appealing starting point for atomistic simulations, but their accuracy for material-specific observables often remains limited without additional reference data (fine-tuning). Here, we systematically quantify how much first-principles data are required to convert universal models into ab initio-accurate material-specific potentials, and ask whether fine-tuning is necessarily preferable to training from scratch. We compare five universal MLIP frameworks, MACE-MP-0, SevenNet-0, GRACE-1L-OAM, MatterSim-v1-5M and ORB-v2, across seven chemically diverse systems incorporating rare and reactive events. Fine-tuning on only 10 AIMD-derived configurations is insufficient for the investigated systems; 200 configurations succeed in favorable cases, but the outcome remains strongly system-dependent. By contrast, 2000 AIMD configurations constitute a robust default, yielding low force and energy errors and reproducing the target material-specific observables. Moderately dense sub-sampling of the AIMD trajectory reduces the required trajectory length tenfold with little loss in model quality. Training from scratch on the same datasets is competitive with, and often slightly more accurate than, naive fine-tuning for MACE and SevenNet, whereas GRACE requires more data. The energy profile for a sulfur-vacancy jump in MoS$_2$ reveals that low trajectory-level errors do not guarantee a correct reaction profile, highlighting the need for observable-level validation. Finally, we show that averaging independently trained models improves predictions in scarce-data regimes at no additional first-principles cost. Together, these results provide practical guidelines for converting limited AIMD reference data into reliable material-specific MLIPs for nanosecond-timescale simulations at near-DFT accuracy.
A correctly solved crystal structure should agree with the experimental data, and its geometry should correspond to a local minimum on the potential energy surface (PES). The idea of verifying crystal structure solutions by comparing them with their geometry-optimized versions was introduced 15 years ago. Recent developments in machine learning interatomic potentials (MLIPs) have made it possible to replace computationally expensive density functional theory (DFT) calculations with AI/neural-network-based alternatives. MLIPs can reach DFT-comparable precision with a substantial gain in speed. We selected one promising MLIP, Universal Models for Atoms, trained on the Open Molecular Crystals 2025 dataset, and processed a prefiltered subset of 216 919 structures from the Cambridge Structural Database. Due to the limitations of the MLIP available when this study commenced, ionic compounds, salts and metal-containing structures were excluded. The current methodology cannot process disordered structures, and available computational resources limit the maximum unit-cell volume that can be treated to 4000 Å3. All structures in the dataset were geometry optimized using the MLIP, and similarity descriptors were calculated to quantify the differences between the original and optimized structures. Automatic analysis was followed by the manual identification of issues indicated by the descriptors' values. We detected anomalies in experimental structures that had already passed all prior validation, as well as limitations in the reliability of the MLIP PES calculations. For 1867 crystal structures, bond-pattern change was observed, while 3331 structures showed a root-mean-square Cartesian displacement greater than 0.25 Å. Future improvements to the methodology and extension to systems not covered by this study are discussed.
A machine learning approach is presented that accelerates DFTB simulations by predicting optimal initial atomic charges and demonstrates that ML-predicted initial charges consistently and significantly improve SCC convergence across diverse chemical systems including organic molecules, biomolecules, water clusters, transition metal oxides and solid electrolytes.
Maximilian L. Ach, Karsten Reuter, C. Panosetti· 0 citations
Coupled-cluster theory defines the accuracy standard for molecular electronic-structure properties but scales too steeply for routine application, whereas density-functional theory is affordable yet systematically biased. We resolve this trade-off with a single equivariant network, MEHnet-MG, that predicts an effective one-electron Hamiltonian from one inexpensive B3LYP/def2-SVP calculation and derives a broad suite of properties from it (energy, optical gap, dipole, quadrupole, polarizability, Mulliken atomic charges, and Mayer bond orders) at coupled-cluster accuracy across nine main-group elements, including the under-served phosphorus, sulfur, and chlorine chemistries. The model is trained on a new in-house dataset of multi-property labels computed at the CCSD(T) level for all nine elements. On a held-out test set, it reduces the error of every property by a factor of 3.8 to 230 relative to semi-local, hybrid, and double-hybrid DFT (referenced to composite CCSD(T)/cc-pVTZ; Methods), while adding only ~25 ms wall time per molecule, delivering coupled-cluster-quality predictions at the cost of a single DFT calculation. Critically, deriving every property from a predicted Hamiltonian rather than pooling per-atom features builds the correct size-scaling into the model architecture: on pi-conjugated oligothiophenes it matches finite-field CCSD polarizability and the EOM-CCSD optical gap to ~2% at the largest sizes where those references remain affordable (44 and 37 atoms, where a single CCSD field point already costs ~500x the model's entire inference) and extrapolates the corrected trends to 58-atom chains, a regime where pooling-based architectures fail by construction. Accurate extrapolation is therefore set by the model's inductive bias rather than by the training data.
