Roman I. Zubatyuk

Macrocyclic peptides are among the most promising therapeutic modalities, yet no general-purpose machine-learned interatomic potential (MLIP) has been systematically benchmarked for their conformational analysis. We show that on the MPCONF196 macrocyclic benchmark, the four-model AIMNet2 ensemble outperforms MACE-OFF23(S), despite macrocycles being entirely absent from its training data. A single AIMNet2 model is 3- to 5-fold faster and needs 5- to 63-fold less GPU memory than the MACE-OFF23 models at their default float64-precision. The four-model ensemble is about as fast as MACE-OFF23 but retains a 3- to 32-fold memory advantage even against MACE-OFF23 in float32. The larger MACE-OFF23(M) and MACE-OFF23(L) models remain more accurate. To narrow this gap, we introduce a continual training protocol that requires new quantum-chemical labels for fewer than 25,000 domain-specific conformers (<0.13% of the training corpus) while preserving conformational-energy accuracy on nonmacrocyclic molecules. On a sequence-disjoint holdout of unseen CREMP macrocyclic peptides, the resulting model, AIMNet2-macro, reduces conformational energy RMSE by 25 to 27% and DFT force RMSE by 20% (25% on backbone atoms) and improves lowest-energy conformer recovery relative to the original AIMNet2. It retains the conformational-energy accuracy of the original model on the nonmacrocyclic drug-like molecules of CONFORGE and also improves accuracy on the structurally distinct macrocycles of MPCONF196, although this smaller improvement is not statistically significant. Among the ten training strategies compared, only continual training on the combined corpus (AIMNet2-macro) achieves this improvement without degrading small-molecule accuracy. AIMNet2-macro is therefore the best balanced, though not the most accurate, of the models tested. As an independent case study, AIMNet2-macro reproduces DFT conformational energies of six cyclosporine A conformers with an RMSE of 0.78 kcal mol–1 and the correct rank ordering, the lowest error among the MLIPs and semiempirical methods evaluated on this system. The continual training strategy has been adopted in the AIMNet2 (2025) release, which unifies small-molecule, macrocyclic, and intermolecular-interaction treatment in a single production potential.

Mechanistic modeling of chemical transformations offers a compelling basis for understanding reactivity and allows for prediction of reaction outcomes before attempting experiments. Despite progress in machine-learned interatomic potentials (MLIPs), we demonstrate that available models lack the accuracy for diverse reaction modeling. With this motivation, we developed a general MLIP for mechanistic modeling of closed-shell carbon, hydrogen, nitrogen, and oxygen reactions, AIMNet2-rxn, using a dataset of ∼4.7 × 10 ⁶ range-separated density functional theory calculations. AIMNet2-rxn enables reaction modeling ∼10 ⁶ faster than the reference quantum mechanical (QM) methods while substantially outperforming graph-based ML, reaffirming the value using three-dimensional chemical information for training. On a test suite of well-known reaction mechanisms—such as amide formation, proton transfers, and pericyclics—AIMNet2-rxn yields 1 to 2 kilocalories per mole accuracy across reaction coordinates without retraining or system-specific fine-tuning. To exploit graphics processing unit parallelism and AIMNet2-rxn efficiency, we introduce a batched nudged elastic band procedure that readily achieves minimum energy pathway search on a millions-of-reactions scale. To demonstrate complex reaction characterization, the thermodynamics of an 11-step pathway producing hydroxymethylfurfural, the experimentally observed major product of glucose pyrolysis, is evaluated. Overall, the accuracy and efficiency afforded by AIMNet2-rxn create opportunities in high-throughput reaction discovery and deep reaction network analysis that would be infeasible with QM methods.

Open‐shell systems such as radical intermediates are central to radical polymerization (RP), combustion, catalysis, and many other chemical and industrial processes, yet their accurate modeling presents significant computational challenges. Most of the current machine learning interatomic potentials do not distinguish between different spin states, making them unsuitable for open‐shell reactive chemistry. Here we present AIMNet2‐NSE (neural spin‐charge equilibration), a neural network potential that incorporates spin‐charge equilibration for accurate treatment of molecules and reactions with arbitrary charge and spin multiplicities. Built upon the AIMNet2 framework, AIMNet2‐NSE is trained on an extensive dataset comprising 20 million closed‐shell neutral and charged molecules, 13 million open‐shell radical configurations, and 200K radical reaction profiles. With explicit handling of spin charges, AIMNet2‐NSE enables prediction of spin‐resolved properties with near‐DFT accuracy while maintaining a favorable linear scaling compared to the polynomial scaling of electronic structure methods. The predictive capabilities and generalizability of our model are confirmed by evaluations on large‐scale radical test sets, the industrially relevant BASChem19 benchmark, and RP reactions. Overall, AIMNet2‐NSE represents a significant advancement in machine learning interatomic potentials, allowing efficient exploration of complex open‐shell systems, and significantly advancing our ability to model radical reaction pathways and reactive intermediates in chemical processes where traditional quantum mechanical methods are computationally prohibitive.

Open‐shell systems such as radical intermediates are central to radical polymerization (RP), combustion, catalysis, and many other chemical and industrial processes, yet their accurate modeling presents significant computational challenges. Most of the current machine learning interatomic potentials do not distinguish between different spin states, making them unsuitable for open‐shell reactive chemistry. Here we present AIMNet2‐NSE (neural spin‐charge equilibration), a neural network potential that incorporates spin‐charge equilibration for accurate treatment of molecules and reactions with arbitrary charge and spin multiplicities. Built upon the AIMNet2 framework, AIMNet2‐NSE is trained on an extensive dataset comprising 20 million closed‐shell neutral and charged molecules, 13 million open‐shell radical configurations, and 200K radical reaction profiles. With explicit handling of spin charges, AIMNet2‐NSE enables prediction of spin‐resolved properties with near‐DFT accuracy while maintaining a favorable linear scaling compared to the polynomial scaling of electronic structure methods. The predictive capabilities and generalizability of our model are confirmed by evaluations on large‐scale radical test sets, the industrially relevant BASChem19 benchmark, and RP reactions. Overall, AIMNet2‐NSE represents a significant advancement in machine learning interatomic potentials, allowing efficient exploration of complex open‐shell systems, and significantly advancing our ability to model radical reaction pathways and reactive intermediates in chemical processes where traditional quantum mechanical methods are computationally prohibitive.

The AIQM series methods are successful neural network-based models that target coupled-cluster accuracy while maintaining high robustness and transferability across various tasks by leveraging Δ-learning. However, the previous AIQM1 and AIQM2 models are limited to molecular systems with four elements: H, C, N, and O, which falls short of meeting the common needs for atomistic simulations. Here, we introduce the extension—AIQM3—that covers three additional chemical elements: S, F, Cl, and approaches coupled cluster level at the speed of a semi-empirical method. AIQM3 maintains the accuracy of its predecessor AIQM2, surpasses the commonly used density functional theory (DFT) method in different types of molecular interactions, and its efficiency is competitive with that of machine learning interatomic potentials on commodity CPU hardware. AIQM3 superiority is showcased for reaction simulations and tasks related to drug design, where it delivers accurate torsion profiles for various real-world drug-like molecules. In addition, AIQM3 can be used for infrared (IR) spectra calculations at a low cost. We provide a web service for the AIQM3 calculations on the Aitomistic Hub at aitomistic.xyz, to democratize and facilitate its use with the assistance of AI agents.

Identifying thermodynamically stable crystal structures remains a key challenge in materials chemistry. Computational crystal structure prediction (CSP) workflows typically rank candidate structures by lattice energy to assess relative stability. Approaches using self-consistent first-principles calculations become prohibitively expensive, especially when millions of energy evaluations are required for complex molecular systems with many atoms per unit cell. Here, we provide a detailed analysis of our methodology and results from the seventh blind test of crystal structure prediction organized by the Cambridge Crystallographic Data Centre (CCDC). We present an approach that significantly accelerates CSP by training target-specific machine-learned interatomic potentials (MLIPs). AIMNet2 MLIPs are trained on density functional theory (DFT) calculations of molecular clusters, herein referred to as n-mers. We demonstrate that potentials trained on gas phase dispersion-corrected DFT reference data of n-mers successfully extend to crystalline environments, accurately characterizing the CSP landscape and correctly ranking structures by relative stability. Our methodology effectively captures the underlying physics of thermodynamic crystal stability using only molecular cluster data, avoiding the need for expensive periodic calculations. The performance of target-specific AIMNet2 interatomic potentials is illustrated across diverse chemical systems relevant to pharmaceutical, optoelectronic, and agrochemical applications, demonstrating their promise as efficient alternatives to full DFT calculations for routine CSP tasks.

Identifying thermodynamically stable crystal structures remains a key challenge in materials chemistry. Computational crystal structure prediction (CSP) workflows typically rank candidate structures by lattice energy to assess relative stability. Approaches using self-consistent first-principles calculations become prohibitively expensive, especially when millions of energy evaluations are required for complex molecular systems with many atoms per unit cell. Here, we provide a detailed analysis of our methodology and results from the seventh blind test of crystal structure prediction organized by the Cambridge Crystallographic Data Centre (CCDC). We present an approach that significantly accelerates CSP by training target-specific machine learned interatomic potentials (MLIPs). AIMNet2 MLIPs are trained on density functional theory (DFT) calculations of molecular clusters, herein referred to as n-mers. We demonstrate that potentials trained on gas phase dispersion-corrected DFT reference data of n-mers successfully extend to crystalline environments, accurately characterizing the CSP landscape and correctly ranking structures by relative stability. Our methodology effectively captures the underlying physics of thermodynamic crystal stability using only molecular cluster data, avoiding the need for expensive periodic calculations. The performance of target-specific AIMNet2 interatomic potentials is illustrated across diverse chemical systems relevant to pharmaceutical, optoelectronic, and agrochemical applications, demonstrating their promise as efficient alternatives to full DFT calculations for routine CSP tasks.

Mechanistic modeling of chemical transformations offers a compelling basis for understanding reactivity and allows for prediction of reaction outcomes before attempting experiments. Despite progress in machine learned interatomic potentials (MLIPs), we demonstrate that available models lack the accuracy for diverse reaction modeling. With this motivation, we developed a general MLIP for mechanistic modeling of organics, AIMNet2-rxn, using a dataset of ~4.7 x 106 range-separated DFT calculations. AIMNet2-rxn enables reaction modeling ~106 faster than the reference quantum mechanical (QM) methods while significantly outperforming graph-based ML, reaffirming the value using 3D chemical information for training. On a test suite of well-known reaction mechanisms—such as amide formation, proton transfers, and pericyclics—AIMNet2-rxn yields 1-2 kcal mol-1 accuracy across reaction coordinates without retraining or system-specific fine-tuning. To exploit GPU parallelism and AIMNet2-rxn efficiency, we introduce a batched nudged elastic band (BNEB) method that readily achieves minimum energy pathway search on a millions-of-reactions scale. To demonstrate complex reaction characterization, the thermodynamics of an 11-step pathway producing hydroxymethylfurfural, the experimentally observed major product of glucose pyrolysis, is evaluated. Overall, the accuracy and efficiency afforded by AIMNet2-rxn creates opportunities in high-throughput reaction discovery and deep reaction network analysis that would be infeasible with QM methods.

Reactive potentials serve as essential tools for investigating chemical reactions with moderate computational costs. However, traditional reactive potentials often depend on fixed, semiempirical parameters, which limits their accuracy and transferability. Overcoming these limitations can significantly expand the applicability of reactive potentials, enabling the simulation of a broader range of reactions under diverse conditions and the prediction of reaction properties, such as barrier heights. This work introduces ANI-1xBB, a novel ANI-based reactive ML potential trained on off-equilibrium molecular conformers generated through an automated bond-breaking workflow. ANI-1xBB significantly enhances the prediction of reaction energetics, barrier heights, and bond dissociation energies, surpassing those of conventional ANI models. Our results show that ANI-1xBB improves transition state modeling and reaction pathway prediction while generalizing effectively to pericyclic reactions and radical-driven processes. Furthermore, the automated data generation strategy supports the efficient construction of large-scale, high-quality reactive data sets, reducing reliance on expensive QM calculations. This work highlights ANI-1xBB as a practical model for accelerating the development of reactive machine learning potentials, offering new opportunities for modeling reaction phenomena.

Reactive potentials serve as essential tools for investigating chemical reactions with moderate computational costs. However, traditional reactive potentials often depend on fixed, semi-empirical parameters, which limits their accuracy and transferability. Overcoming these limitations can significantly expand the applicability of reactive potentials, enabling the simulation of a broader range of reactions under diverse conditions and the prediction of reaction properties, such as barrier heights. This work introduces ANI-1xBB, a novel ANI-based reactive ML potential trained on off-equilibrium molecular conformers generated through an automated bond-breaking workflow. ANI-1xBB significantly enhances the prediction of reaction energetics, barrier heights, and bond dissociation energies, surpassing conventional ANI models. Our results show that ANI-1xBB improves transition state modeling and reaction pathway prediction while generalizing effectively to pericyclic reactions and radical-driven processes. Furthermore, the automated data generation strategy supports the efficient construction of large-scale, high-quality reactive datasets, reducing reliance on expensive QM calculations. This work highlights ANI-1xBB as a practical model for accelerating the development of reactive machine learning potentials, offering new opportunities for modeling reaction phenomena.