Reproduce “Inverse design of bespoke interatomic potentials via active learning by information-matching” on OpenResearch

About this paper

Yonatan Kurniawan, Logan D. Williams, Amit Samanta, Ilia Nikiforov, Daniel Schwalbe-Koda, Mark K. Transtrum, Ellad B. Tadmor, Vincenzo Lordi and 1 more

This work introduces the Active Learning by Information-Matching (ALIM) framework to inversely design bespoke interatomic potentials (IPs) for specific material properties. The framework efficiently selects minimal training data to achieve prescribed uncertainty targets for quantities of interest, demonstrating its application to predict tantalum's plastic strength via indicator properties and revealing the impact of model form limitations on prediction uncertainty.

Key insights

• ALIM efficiently identifies highly informative training configurations, with only a small fraction (0.5-1.0%) of candidate environments being sufficient to constrain IPs for targeted indicator properties.
• Model error, stemming from the functional form of the IP, significantly impacts prediction uncertainty, leading to overconfident initial predictions even when parameter uncertainty is minimized.
• The indirect approach of using indicator properties is viable for complex QoIs, but robust uncertainty quantification for the ultimate target relies critically on the model's flexibility and the sufficiency of the indicator property set.

Summary by alphaXiv.