JAXRENS moves nested sampling onto GPUs: a ZnS phase diagram with its melting line finishes in 5.8 days on four H100s
Synopsis
Nested sampling can deliver thermodynamics across a whole temperature range from a single run, but combining it with modern machine-learned interatomic potentials (MLIPs) for finite-temperature phase-stability prediction has been limited by computational throughput. JAXRENS treats the whole population of sampling walkers as one batched computation, compiled into a single GPU kernel through JAX, with multi-GPU sharding and replica exchange. Starting from a MACE foundation model and distilling an active-learning NeuralIL potential, a 32-atom ZnS pressure-temperature phase diagram, including a melting line not measured under pressure, completed in 5.8 days on four H100s, against an extrapolated 2.2 years for a serial ASE-based driver.
Interpretation
Finite-temperature phase-stability prediction can now be driven directly by machine-learned interatomic potentials without first assuming the competing phases. JAXRENS merges the independent walker evolutions of nested sampling into one batched computation, so the sampler couples directly to modern potentials such as MACE, NeuralIL and Nequix. Earlier atomistic nested-sampling codes, pymatnest and FreeBird.jl, were built for CPUs; through their ASE interface they evaluate one configuration at a time on a GPU, giving up the walker parallelism that MPI provided. On the Si32 benchmark at 27 Å3/atom, JAXRENS costs 0.14 ms per model call at batch width 64; pymatnest on sixteen CPU cores costs 4.5 ms, and the serial GPU paths cost 5.2 ms (FreeBird.jl) and 6.6 ms (pymatnest).
A single nested-sampling run yields free energies and heat capacities over the whole temperature range, with phase transitions appearing as heat-capacity peaks and no order parameter prescribed. Replica exchange lets simulations at different pressures swap configurations, improving ergodicity on complex energy landscapes. Conventional temperature-based sampling locates these regions only if the relevant phases are known and targeted in advance, whereas nested sampling scans configuration space from high to low enthalpy. In the final 32-atom ZnS run, 30–77% of attempted swaps are accepted per neighbouring pressure pair, and the B3–B1 transition appears as a diagonal of reduced acceptance following the boundary.
Starting from a foundation model and distilling a system-specific potential by active learning reduces the system-specific input of a first-principles-quality phase diagram to a limited number of reference calculations. In the 16-atom foundation-model cell the melting peak is broad and the solid-solid boundary is resolved only on the pressure grid, so larger cells and denser pressure grids are needed. 1000 configurations strided from the foundation-model trajectories were labelled with r2SCAN to train a NeuralIL committee; four active-learning iterations labelled 1792 configurations in total, and the final run used 64 pressures on a 32-atom cell, completing in 5.8 days on four H100s.
Perspective
The demonstrated scope is Si32 benchmark timing and a ZnS phase diagram computed on a 32-atom cell at 64 pressures between 0 and the stated upper pressure; the ambient-pressure melting temperature of 2031 K lies between the measured melting points of stoichiometric and Zn-rich ZnS. The intended reader is a materials or computational-physics researcher working on atomistic thermodynamic sampling.
The authors note that the absolute position of the B3–B1 transition is inherited from the r2SCAN transition pressure rather than from the sampling, and that this value lies above the extrapolation of the high-temperature X-ray data. The ambient-pressure melting point is itself disputed in the literature, which the authors attribute to composition differences.
