Balancing accuracy and computational cost in computational chemistry

About

As molecular simulations increase in scale, the environmental cost of computational chemistry is becoming harder to ignore. High‑accuracy quantum methods remain too expensive for routine large‑scale use, while empirical force fields trade fidelity for speed. Machine learning potentials (MLPs) offer a middle ground, yet their computational and energy demands remain significant, especially for long timescales and large systems. This reflects a broader challenge in responsible research computing: balancing accuracy, computational efficiency, and environmental impact. We present a strategy to reduce the computational footprint of MLPs while preserving accuracy and stability. Our active‑learning framework minimizes evaluation cost and training‑set size by using kernel density estimation on node energies to identify out‑of‑distribution configurations, avoiding costly ensemble‑based uncertainty quantification. Benchmarking against reactive force fields and foundation MLPs assesses accuracy, speed, memory use and environmental cost. Applied to a hydrated silicate ionic liquid, our tailored MLP delivers major speedups, lower memory use, and reduced energy per simulated picosecond, enabling more sustainable molecular simulations.

00:00:00 Slide 1

00:00:28 Slide 2

00:00:58 Slide 3

00:01:42 Slide 4

00:03:02 Slide 5

00:03:59 Slide 6

00:06:00 Slide 7

00:07:29 Slide 8

00:08:08 Slide 9

00:08:33 Slide 10

00:09:20 Slide 11