MACE-POLAR-1: A Polarisable Electrostatic Foundation Model for Molecular Chemistry
Ilyes Batatia, William J. Baldwin, Domantas Kuryla, Joseph Hart, Elliott Kasoar, Alin M. Elena, Harry Moore, Mikołaj J. Gawkowski
Abstract
Accurate modelling of electrostatic interactions and charge transfer is fundamental to computational chemistry, yet most machine learning interatomic potentials (MLIPs) rely on local atomic descriptors that cannot capture long-range electrostatic effects. We present a new electrostatic foundation model for molecular chemistry that extends the MACE architecture with explicit treatment of long-range interactions and electrostatic induction. Our approach combines local many-body geometric features with a non-self-consistent field formalism that updates learnable charge and spin densities through polarisable iterations to model induction, followed by global charge equilibration via learnable Fukui functions to control total charge and total spin. This design enables an accurate and physical description of systems with varying charge and spin states while maintaining computational efficiency. Trained on the OMol25 dataset of 100 million hybrid DFT calculations, our models achieve chemical accuracy across diverse benchmarks, with accuracy competitive with hybrid DFT on thermochemistry, reaction barriers, conformational energies, and transition metal complexes. Notably, we demonstrate that the inclusion of long-range electrostatics leads to a large improvement in the description of non-covalent interactions and supramolecular complexes over non-electrostatic models, including sub-kcal/mol prediction of molecular crystal formation energy in the X23-DMC dataset and a fourfold improvement over short-ranged models on protein-ligand interactions. The model's ability to handle variable charge and spin states, respond to external fields, provide interpretable spin-resolved charge densities, and maintain accuracy from small molecules to protein-ligand complexes positions it as a versatile tool for computational molecular chemistry and drug discovery.
AI Impact Assessments
(3 models)Scientific Impact Assessment: MACE-POLAR-1
1. Core Contribution
MACE-POLAR-1 introduces a physics-informed electrostatic extension to the MACE machine learning interatomic potential (MLIP) architecture, addressing a fundamental limitation of local message-passing neural networks: their inability to capture long-range electrostatic interactions. The key innovation is a non-self-consistent field formalism that iteratively updates learnable spin-resolved charge densities through polarizable iterations, combined with Fukui-function-based global charge equilibration to enforce total charge and spin constraints. This avoids the well-known pathologies of classical charge equilibration (QEq) schemes—incorrect fractional charge dissociation, cubic polarizability scaling—while capturing polarization and long-range charge transfer without the cost of full self-consistency.
The energy decomposition into local (MACE), electrostatic (smeared Coulomb), and non-local (learned correction) terms provides a principled separation where short-range quantum effects are handled by flexible neural networks while long-range interactions are constrained to physically motivated functional forms. This is crucial for extrapolation from small training systems to large biomolecular complexes.
2. Methodological Rigor
The theoretical framework is well-grounded, with clear connections to conceptual DFT (Fukui functions, chemical potential) and electronic structure theory (spin-resolved density, Gaussian multipole expansions). The supplementary material provides a rigorous derivation connecting the Fukui equilibration to conceptual DFT, lending theoretical legitimacy to what could otherwise appear as an ad hoc normalization.
The benchmark campaign is exceptionally comprehensive—spanning GSCDB138 thermochemistry, reaction barriers, non-covalent interactions (S22, S66, IHB100x10), supramolecular complexes (S30L), protein-ligand interactions (QUID, PLF547, PLA15), molecular crystals (X23-DMC, CPOSS209), transition metal complexes, redox potentials, liquid densities, and external field response. The inclusion of close ablation baselines (MACE-OMOL with identical local architecture but no electrostatics) enables clean attribution of improvements to the electrostatic component. Comparison against multiple state-of-the-art models (UMA, OrbMol, SO3LR, AIMNet2-NSE, FENNIX-BIO-2) is thorough.
However, several methodological concerns arise: (1) no training on partial charges or dipoles was performed, and the authors note dipole data were unavailable—this is a missed opportunity for validation; (2) the systematic ~5-10% overestimation of liquid densities across all models is attributed to the ωB97M-V functional, but model-architecture contributions are not rigorously disentangled; (3) the redox potential results require per-model constant shifts, limiting interpretation of absolute accuracy.
3. Potential Impact
The practical impact is substantial across multiple domains:
The architectural innovation—polarizable field updates with Fukui equilibration—provides a transferable blueprint for other MLIP architectures. The explicit demonstration that physics-constrained long-range terms outperform flexible message-passing at equivalent or greater range (UMA-M-1P1 at 66 Å performs worse than MACE-POLAR-1-L at 18 Å on PLA15) is an important finding for the field.
4. Timeliness & Relevance
This work arrives at a critical juncture. The release of OMol25 (100M hybrid DFT calculations) has enabled a new generation of molecular foundation models, but all prior OMol-trained models (UMA, MACE-OMOL, OrbMol) lack explicit electrostatics. The paper convincingly demonstrates that this omission becomes catastrophic for large biomolecular systems, charged species, and condensed-phase extrapolation—precisely the applications most relevant to drug discovery and materials science. The failure modes demonstrated (Cl₂ dissociation in presence of distant Cl⁻, charge delocalization in solvated Fe²⁺/Fe³⁺) are not merely academic but represent real obstacles to deploying MLIPs in production settings.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Missing Comparisons: No comparison to classical polarizable force fields (AMOEBA, Drude) on biomolecular benchmarks, which would contextualize the practical advantage for the target drug discovery audience.
Overall Assessment
This paper represents a significant methodological advance in molecular MLIPs, successfully integrating physics-based long-range electrostatics into a foundation model framework. The improvements on challenging benchmarks are convincing and practically relevant. The work sets a new standard for what molecular foundation models should include and provides a clear path for the field. The main limitation is that some quantitative aspects (polarizabilities, liquid densities) reveal room for improvement, and computational cost analysis is needed to assess practical deployment.
Generated Apr 19, 2026
Comparison History (188)
Paper 1 has higher impact potential due to its combination of major-scale training (100M hybrid DFT calculations), a broadly useful foundation model for energies/forces with explicit long-range electrostatics, induction, and charge/spin-state handling, and strong benchmark evidence spanning thermochemistry, barriers, noncovalent interactions, crystals, transition metals, and protein–ligand systems. Its applications extend across computational chemistry, materials, and drug discovery. Paper 2 is novel and practical for complex generative sampling, but appears narrower in scope and validation, and is less likely to become a general-purpose workhorse than a high-accuracy electrostatic MLIP foundation model.
Paper 1 likely has higher impact due to its combination of (i) major scale (trained on 100M hybrid-DFT calculations), (ii) broad applicability from small molecules to protein–ligand complexes and crystals, and (iii) addressing a central limitation of ML interatomic potentials via explicit long-range electrostatics, induction, and charge/spin control. The reported benchmark breadth (thermochemistry, barriers, noncovalent, transition metals, protein–ligand) and strong gains in supramolecular and biomolecular settings suggest wide cross-field relevance and near-term real-world use (drug discovery, materials). Paper 2 is innovative and rigorous but currently narrower in scope/datasets.
Paper 2 likely has higher impact due to a more novel, broadly enabling foundation-model contribution: explicit long-range electrostatics and induction with variable charge/spin, trained at massive scale (100M hybrid-DFT). Its applicability spans thermochemistry, barriers, noncovalent interactions, crystals, transition metals, and protein–ligand systems, positioning it for wide adoption in molecular chemistry and drug discovery. Paper 1 is rigorous and practically valuable (distillation for cheaper high-accuracy condensed-phase sampling), but is more of a performance/efficiency advance on existing MLIP paradigms with narrower domain focus.
Paper 2 addresses a fundamental limitation in current MLIPs by explicitly modeling long-range electrostatics and variable charge/spin states. Its massive scale (trained on 100M configurations) and demonstrated success in complex applications like protein-ligand interactions and molecular crystals suggest transformative potential for computational chemistry and drug discovery. While Paper 1 offers an elegant methodological improvement for sample efficiency, Paper 2's comprehensive foundation model presents a broader and more immediately impactful contribution across multiple disciplines.
MACE-POLAR-1 represents a major methodological advance in machine learning interatomic potentials by solving the long-standing challenge of incorporating long-range electrostatics and polarization into foundation models. Trained on 100M DFT calculations, it achieves chemical accuracy across diverse benchmarks and demonstrates broad applicability from small molecules to protein-ligand complexes and drug discovery. Its breadth of impact across computational chemistry, materials science, and drug discovery, combined with the foundational nature of the model architecture, gives it substantially higher potential impact than Paper 1, which, while scientifically interesting, addresses a more specialized topic in confined ionic transport.
Paper 1 presents a foundation model with broad, immediate applicability in drug discovery and materials science. By successfully integrating long-range electrostatics into machine learning interatomic potentials and training on a massive dataset, it solves a major limitation in current MLIPs, achieving breakthroughs in protein-ligand modeling. While Paper 2 offers a highly valuable algorithmic memory reduction for specific quantum chemistry workflows, Paper 1 promises much wider cross-disciplinary impact, broader real-world applications, and represents a more significant paradigm shift in scalable computational chemistry.
MACE-POLAR-1 represents a foundational advance in ML interatomic potentials by solving the long-standing challenge of incorporating long-range electrostatics and polarization into MLIPs at scale. Trained on 100M DFT calculations, it achieves chemical accuracy across diverse benchmarks from small molecules to protein-ligand complexes, with broad applicability to drug discovery, materials science, and molecular chemistry. Paper 2, while methodologically sound and valuable for reactive simulations, addresses a more specialized niche (active learning for transition states in electrochemistry). Paper 1's breadth of impact, foundation model nature, and versatility across charge/spin states give it substantially higher potential impact.
MACE-POLAR-1 represents a major advance in machine learning interatomic potentials by incorporating long-range electrostatics and polarization into a foundation model trained on 100M DFT calculations. It addresses a fundamental limitation of current MLIPs, demonstrates broad applicability from small molecules to protein-ligand complexes, and achieves chemical accuracy across diverse benchmarks. Its impact spans computational chemistry, drug discovery, and materials science. Paper 2, while novel in exploring Pauli principle effects in polaritonic chemistry, addresses a more niche topic with narrower immediate applications and smaller community impact.
Paper 1 has higher impact potential due to its foundation-model scale (100M hybrid-DFT training set), broad applicability (from small molecules to protein–ligand and crystals), and a novel integration of long-range electrostatics, polarizable iterations, and charge/spin control that addresses key MLIP limitations. Its demonstrated improvements on noncovalent interactions and variable charge/spin systems are directly relevant to materials, catalysis, and drug discovery. Paper 2 is methodologically elegant and useful for SCF acceleration, but is validated on a small set of closed-shell systems, suggesting narrower and earlier-stage impact.
Paper 1 presents a foundation model trained on a massive dataset (100M calculations) that addresses a major limitation in machine learning interatomic potentials by incorporating long-range electrostatics. Its proven ability to handle complex systems like protein-ligand interactions and molecular crystals gives it broad applicability in drug discovery and materials science. In contrast, Paper 2 focuses on accelerating SCF convergence for DFT calculations, which, while useful, represents a more incremental methodological improvement tested primarily on small, simple molecules.