BLAST

Bridging Length/timescales via Atomistic Simulation Toolkit

BLAST (Bridging Length/timescales via Atomistic Simulation Toolkit) is a suite of tree-based and reinforcement-learning optimization methods for developing empirical and neural-network interatomic potentials from sparse first-principles datasets. The methodology — originally inspired by AI strategies for board games — treats force-field construction as a sequential decision-making problem with a continuous action space, dramatically accelerating the search for accurate and transferable potentials.

BLAST has been used to develop potentials for a broad family of materials: silica, silicene, arsenene, phosphorene, bismuthene, gold and nickel nanoclusters, transition metals, and high-entropy MXenes. The framework has been extended with multi-reward and hierarchical-RL variants, and most recently with symbolic-regression hybrids that yield physically interpretable potentials.

Status: Active development

Role: Lead developer

Selected method papers

  • Manna et al., Learning in Continuous Action Space for Determination of High Dimensional Potential Energy Surfaces, Nature Communications 13, 368 (2022). DOI
  • Varughese et al., Physically Interpretable Interatomic Potentials via Symbolic Regression and Reinforcement Learning, npj Computational Materials 12, 84 (2026). DOI
  • Koneru et al., Multi-Reward Reinforcement Learning Based Development of Inter-atomic Potential Models for Silica, npj Computational Materials 9, 125 (2023). DOI
  • Loeffler & Manna et al., Active Learning a Neural Network Model for Gold Clusters & Bulk from Sparse First Principles Training Data, ChemCatChem 12, 4796 (2020) — Cover article. DOI

Citation

@article{manna2022blast,
  title   = {Learning in continuous action space for determination of
             high dimensional potential energy surfaces},
  author  = {Manna, Sukriti and Loeffler, Troy D. and Batra, Rohit and
             Banik, Suvo and Chan, Henry and others},
  journal = {Nature Communications},
  volume  = {13},
  pages   = {368},
  year    = {2022},
  doi     = {10.1038/s41467-021-27849-6}
}