Machine Learning Potentials

The Core Challenge

Computational chemistry faces a critical tradeoff: classical force fields are fast but inaccurate for reactive systems and complex chemical interactions, while first-principles quantum mechanical methods provide accuracy but scale poorly with system size. This limitation has historically constrained the scope of computationally accessible chemistry.

Our Solutions: The AIMNet Family

The centerpiece of our work is the AIMNet family of neural network potentials. AIMNet2 demonstrates that neural networks can generalize across diverse chemical domains without retraining, covering 14 elements found in over 90% of drug-like molecules. The framework uniquely integrates ML-parameterized interactions with physics-based long-range electrostatics.

Key Capabilities

• Treatment of neutral molecules, ions, radicals, and reactive intermediates within unified frameworks
• 1-2 kcal/mol accuracy relative to density functional theory
• Computational speedups of several orders of magnitude enabling high-throughput applications
• Chemical transferability as primary design objective

Reactive Chemistry Extension

AIMNet2-rxn addresses chemical transformation modeling across bond-breaking and bond-forming regions, achieving roughly 6 orders of magnitude speedup compared to quantum mechanical calculations while maintaining chemical reliability for transition state energetics and reaction barrier predictions.

Methodological Foundations

Our research emphasizes:

• Chemical Transferability: Primary design objective ensuring models work across chemical domains
• Physically Informed Representations: Respecting molecular symmetry and fundamental physics
• Robust Extrapolation: Reliable behavior for charged and reactive species
• Seamless Integration: Compatible with molecular dynamics engines and discovery pipelines

Deployment and Impact

The original ANI publication has accumulated over 1,500 citations. Our methods are deployed at pharmaceutical and chemical companies including Dow, BASF, GSK, Pfizer, and Genentech, accelerating drug discovery and materials design.

Software Tools

Our neural network potentials are available through multiple open-source packages:

• AIMNet2: State-of-the-art neural network potential calculator
• Auto3D: Automated conformer generation from molecular structures
• TorchANI: PyTorch implementation of ANI potentials with extensive documentation
• aimnetcentral: Fast ML interatomic potential for molecular dynamics simulations

Future Directions

Emerging priorities include:

• Unified potentials for photochemistry and excited states
• Advanced catalysis modeling with metal centers
• Adaptive learning protocols for autonomous discovery
• Uncertainty-quantified predictions to guide high-throughput screening
• Integration with automated experimentation platforms

Collaborations

This work is carried out in collaboration with experimental groups at CMU, national laboratories (ORNL, LANL), pharmaceutical companies, and academic partners worldwide.

Publications

See our publications page for the latest research papers in this area, including the foundational ANI and AIMNet publications.