cusp-ai-oss / kUPS

What is kUPS?

kUPS is a high performance toolkit for atomistic simulations in Python using JAX, letting you build molecular dynamics, Monte Carlo, and geometry optimization workflows that run batched across thousands of systems on CPU, GPU, or TPU. It provides composable samplers, potentials like Lennard-Jones, Ewald Coulomb, and ML force fields, plus CLI apps for quick GCMC adsorption or NVT/NPT MD runs from CIF files. Users get hardware-accelerated, differentiable simulations without rewriting code for accelerators.

Why is it gaining traction?

Unlike rigid C++ codes like LAMMPS, kUPS snaps primitives together freely—mix MD propagators with MC moves or plug in PyTorch ML models via seamless interop—while JIT-compiling everything for high performance Python on GitHub. Batched runs vectorize independent replicas effortlessly, and full differentiability opens ML workflows like force-matching. Devs love the zero-overhead GPU/TPU scaling and YAML-driven CLIs for rapid prototyping.

Who should use this?

Materials scientists screening MOFs/zeolites for CO2 adsorption via GCMC, or computational chemists running high-throughput MD with ML potentials on clusters. JAX users building custom high performance computing GitHub pipelines for equivariant models, or researchers needing end-to-end differentiable sims from CIF to observables.

Verdict

Solid pick for JAX-native atomistic work—excellent docs, rigorous CI validation, and Apache 2.0—but at 87 stars and 1.0% credibility score, it's early-stage; test on your hardware first. Production-ready for batched prototypes, pair with established tools for long runs.

(198 words)