PARS pins refinement to a frozen prior: Ti vacancy-cluster energy error falls 81%, and a distilled Si/Ge potential matches a density-functional-trained one on interfacial resistance
Synopsis
Foundation interatomic potentials carry systematic task-specific bias, yet the local environments that need first-principles labels often sit in cells too large for first-principles calculation. PARS fixes the acquisition criterion on a frozen foundation representation, scores each atomic environment against element-conditioned pretraining prototypes, and rebuilds selected environments as compact periodic cells for labelling and iterative fine-tuning of the student potential. On Ti vacancy clusters the per-vacancy mean absolute error fell 81% relative to the pretrained model; for Si/Ge, 800 labelled cells of 10–128 atoms sufficed for a distilled potential whose interfacial thermal resistance in a 362,880-atom simulation matched one trained directly on density-functional data.
Interpretation
Refining a foundation potential no longer relies on model uncertainty that shifts during training; a frozen pretrained representation decides which atomic environments lack support. Earlier active-learning criteria came from the current model or an ensemble, such as extrapolation grades, ensemble force disagreement, Gaussian-process variance or latent-space distance, and these criteria move as refinement proceeds. In an 80 GPa silicon trajectory, 269 of 43,416 atom–frame evaluations crossed the monitoring threshold and 41 crossed the strict out-of-distribution boundary, with the first monitoring crossing near 0.35 ps; the same trajectory at 20 GPa produced no crossing in 2,012 evaluations.
Crossing atoms correspond to local structures with raised coordination, not to instantaneous pressure or temperature. Relative to supported atoms in the same frame, crossing events showed a median 13.7% increase in coordination number within 2.6 Å and a 3.3% increase in mean first-to-fourth neighbour distance. The support score's Spearman correlation was 0.65 with coordination number and 0.69 with that neighbour-distance metric, but only 0.02 with instantaneous pressure and with temperature.
Selected atoms are rebuilt as small periodic cells with a fixed core, and only those cells receive first-principles labels instead of the whole parent cell. Reconstruction searches six face extensions of the parent cell for a periodic closure, rejecting closures that break the core, alter the centre's local chemistry or create short periodic contacts, then relaxes the movable buffer with the current student potential. In the Ti test, 60–96-atom cells were labelled from 768-site parent cells; the property models used 756 (NaCl), 384 (NH3/Pt), 192 (Ti) and 128 (LiI/Li2S) first-principles labels.
Refined potentials correct systematic biases of the foundation model on several target properties, and those labels carry to much larger simulation scales. The pretrained potential underestimated, in all three tests, a quantity set by a higher-energy or higher-coordinated configuration: the B1–B2 transition pressure, the NH3 hopping barrier on Pt(111) and the vacancy formation energy. Ti vacancy-cluster per-vacancy mean absolute error fell from 0.568 to 0.108 eV; for Si/Ge, the refined MatterSim labelled 1,347 configurations and the distilled NEP gave a thermal boundary resistance of 4.47 m² K GW⁻¹ in a 362,880-atom, 76.0 nm cell, the same as the CPL potential trained directly on density-functional data.
Perspective
The demonstrated scope is support diagnostics in high-pressure silicon, one silicon transfer benchmark and four property tests (NaCl phase equilibrium, NH3/Pt(111) hopping, hcp Ti vacancy clusters, LiI/Li2S interfacial exchange), plus Si/Ge interfacial thermal resistance from distillation and 362,880-atom non-equilibrium molecular dynamics; the intended reader works on machine-learned interatomic potentials and materials simulation.
The support score measures representation support, not predictive error: environments well covered in pretraining but poorly predicted do not cross the support boundary, and scores can stay high after a model has learned an environment; thresholds are anchored to the starting configurations and need recalibration when the chemistry or thermodynamic range changes. The LiI/Li2S exchange improvement appears only at 333–373 K, with shape error rising at 293 and 313 K, and the refined model carries a long-range term the pretrained model lacks. For Si/Ge the two potentials agree on interfacial resistance but differ by about 20% in bulk conductivity, so agreement on one transport property does not certify others.
