Software

Reciprocal-Space Gated Attention (RSGA)

The core RSGA modules used to introduce a reciprocal-space long-range correction into a short-range local model via gated linear attention.

Repository:  reciprocal_space_gated_attention

GPU-optimized Reciprocal-Space Gated Attention (RSGA)

Standalone RSGA modules for adding a reciprocal-space long-range correction to a short-range atomistic model. Shared geometry, reciprocal-grid caching, and chunked evaluation reduce repeated work and GPU memory use. The implementation was used for the In₂Se₃ molecular-dynamics study below, with strict float64 performance validated on NVIDIA A100 and H100-class GPUs.

Repository:  RSGA_OPTIMIZED

Associated preprint: Boundary-driven phase transformation in layered 3R In₂Se₃ captured by long range machine learning molecular dynamics (ChemRxiv, 2026).

GPU-optimized MACERSGA: RSGA integrated with MACE

The MACE-integrated implementation combines local MACE representations with RSGA corrections at each interaction layer. It provides the optimized float64 code used in the 2026 In₂Se₃ simulations, with shared geometry and reciprocal-grid caching across RSGA layers.

Repository:  RSGA_MACE_OPT

Associated preprint: Boundary-driven phase transformation in layered 3R In₂Se₃ captured by long range machine learning molecular dynamics (ChemRxiv, 2026).

EloKriti: Active learning for BPA-free dental resin candidates

Chemistry-aware active learning and multi-objective Bayesian optimization prioritize candidate backbones for explicit virtual-curing and molecular-dynamics evaluation. The workflow explores trade-offs in elastic modulus, volumetric shrinkage, and fractional free volume; reported properties and Pareto assignments come from the molecular-dynamics evaluations.

Repository:  Elokriti

Associated preprint: Accelerated Discovery of BPA-Free Dental Resin Candidates through Active Learning of Network-Emergent Properties (ChemRxiv, 2026).

Active-learning workflow for Gaussian Approximation Potentials (GAP)

Automated active learning workflow for training Gaussian Approximation Potentials (GAP) for atomistic simulations. The workflow closes the loop between sampling, uncertainty estimation, and model refinement.

Repository:  active-learning-md
Tutorial / hands-on materials:  psik-workshop-AL-GAP

GPU-accelerated active learning with Gaussian Processes / Deep Kernel Learning

A scalable active learning framework implemented in PyTorch/GPyTorch. Uses Gaussian process regression and Deep Kernel Learning on GPUs to accelerate query selection and model improvement.

Repository:  ECG_ActiveLearning

AL4GAP: Ensemble active learning at scale

AL4GAP is an automated workflow for fitting machine-learned interatomic potentials (MLIPs) across combinatorial chemical spaces. It uses Cray SmartSim for ensemble-style active learning on leadership-class high-performance computing systems.

Repository:  AL4GAP_JCP

Catastrophic forgetting mitigation in chemistry-oriented LLMs

This project studies catastrophic forgetting in a Mistral-7B model fine-tuned on sequential chemistry tasks (e.g., reaction yield prediction) and develops continual / life-long learning strategies to preserve prior knowledge while learning new chemistry data.

JAX / Equinox implementation:  CL_MISTRAL7B_REACT

MOLAN: Machine Learning Workflow for Molecular Analysis

MOLAN provides suite of tools for unsupervised, supervised learnings and inverse design models for molecular melting point.

Repository:  molan

AI4PFAS: Deep learning for PFAS toxicity

AI4PFAS provides a suite of tools for molecular machine learning and deep learning to assess toxicity in PFAS-class compounds.

Repository:  AI4PFAS