Ultra-fast interpretable machine-learning potentials

The paper addresses the long-standing speed–accuracy trade-off in atomistic simulations. Electronic-structure methods (e.g., DFT) are accurate but computationally expensive, whereas traditional empirical potentials are fast but limited in accuracy and transferability. Modern machine-learning potentials (MLPs) can approach DFT accuracy but are still orders of magnitude slower than empirical models and often lack interpretability. The research question is whether an interpretable, linear machine-learning potential based on low-order many-body terms and fast-to-evaluate basis functions can achieve accuracy close to state-of-the-art MLPs while matching the speed of the fastest empirical potentials. The authors propose an ultra-fast (UF) potential using effective two- and three-body terms expressed in a cubic B-spline basis and trained via regularized linear regression to deliver near-DFT accuracy at empirical-potential speeds, enabling large systems and long time scales.

The work situates itself among empirical pair and many-body potentials (e.g., Lennard-Jones, Morse, EAM) and modern MLPs. Linear and polynomial-based MLPs include SNAP and qSNAP (bispectrum-based linear and quadratic models), MTP, PACE (atomic cluster expansion), aPIP, and ChIMES (explicit three-body). Speed-focused approaches use fast basis functions and splines; prior efforts project kernel models onto spline bases or employ quadratic spline bases for two- and three-body forms, showing speed gains over MTPs. Polynomial symmetry functions with compact support and spline-based modified EAM potentials have shown strong speed and competitive accuracy despite lower complexity. Benchmarks and datasets from prior work (e.g., GAP on tungsten) provide established baselines for accuracy and cost comparisons.

• Potential form: Truncate the many-body expansion at two- or three-body terms and represent them as functions of pairwise distances: E = sum over pairs V2(r_ij) + sum over triplets V3(r_ij, r_ik, r_jk). For elemental tungsten, species dependence and one-body terms are omitted. Locality is enforced via finite cutoffs (5.5 Å for two-body; 4.25 Å for three-body in UF2,3).
• Basis representation: Use cubic B-spline basis functions (natural cubic splines with compact support) for V2 and tensor-product cubic B-splines for V3. V2(r) = Σ c_n B_n(r); V3(r_ij,r_ik,r_jk) = Σ c_lmn B_l(r_ij) B_m(r_ik) B_n(r_jk). Constraints enforce smooth approach to zero at the upper cutoff and monotonic repulsion at short distances. Due to compact support, at most 4 basis functions contribute per pair and 64 per triplet evaluation, yielding high computational efficiency.
• Forces: Computed analytically via derivatives of spline expansions; derivatives of B-splines are defined recursively, enabling efficient force evaluation.
• Optimization: Fit all spline coefficients jointly using regularized linear least squares (Tikhonov regularization). The loss includes energy and force residuals, with a weighting parameter balancing their contributions after normalization by training-set standard deviations. Regularization includes ridge (λ1) and a second-difference smoothness penalty (λ2) applied along each spline dimension to control coefficient magnitude and curvature, promoting smooth, physically reasonable potentials. The convex problem is solved efficiently (e.g., LU decomposition).
• Basis and knots: Use natural cubic splines with uniform knot spacing; based on convergence tests, 25 uniformly spaced knots were used. Last three coefficients are set to zero to ensure smooth cutoffs. Compact support ensures evaluation cost does not grow with the total number of knots.
• Data and training protocol: A tungsten dataset (9,693 configurations) with diverse configurations (bcc cells, MD snapshots, surfaces, vacancies, gamma surfaces, dislocation quadrupoles) computed with DFT-PBE energies/forces/stresses. Models were typically trained on 20% of data and tested on 80% (RMSE metrics). Comparative models: LJ and Morse (optimized via BFGS), EAM4 (pre-existing), SNAP/qSNAP (MAML), GAP (QUIP), MTP (MLIP).
• Benchmarks and calculations: Accuracy assessed via RMSE on energies, forces, and phonon frequencies; additional derived properties include lattice constant, elastic constants C11, C12, C44, bulk modulus, surface energies (E100, E110, E111), and vacancy formation energy. Phonons computed with Phonopy; elastic constants via Elastic; melting temperature by two-phase LAMMPS simulations. Computational cost measured on a single AMD EPYC 7702 thread; UF2,3 cost reported as bounds based on flop ratios and Python implementation speed ratios.

• Speed–accuracy trade-off: UF potentials (UF2 and UF2,3) are as fast as the fastest empirical potentials (LJ, Morse) while approaching the accuracy of state-of-the-art MLPs (e.g., MTP, GAP). Reported relative computational costs (ms/step, normalized): UF2 ≈ 0.79; UF2,3 ≈ 10.5 (bounds used); Morse ≈ 1.55; EAM4 ≈ 2.92; MTP ≈ 45.7; qSNAP ≈ 145; SNAP ≈ 443; GAP ≈ 3070.
• Retrieval of reference models: UF2 and UF2,3 can exactly reproduce two- and three-body reference potentials (e.g., Lennard-Jones for W, Stillinger–Weber for Si) given sufficient basis resolution and data.
• Accuracy on DFT (tungsten): • UF2 (pair-only) achieves competitive errors: Energy RMSE ≈ 0.027 eV/atom; Force RMSE ≈ 0.387 eV/Å; Phonon RMSE ≈ 0.230 THz. • UF2,3 improves further: Energy RMSE ≈ 0.005 eV/atom; Force RMSE ≈ 0.152 eV/Å; Phonon RMSE ≈ 0.263 THz, approaching MTP and GAP accuracy at far lower cost. • Comparative models: MTP (0.017 eV/atom, 0.146 eV/Å, 0.376 THz), SNAP (0.014, 0.189, 0.270), qSNAP (0.010, 0.167, 0.256), GAP (0.006, 0.169, 0.291), Morse (0.040, 0.480, 1.140), EAM4 (0.088, 0.803, 0.301).
• Derived properties: UF2 yields errors comparable to SNAP/qSNAP for many properties (elastic constants, surfaces, vacancy, phonons) and outperforms LJ/Morse substantially. UF2,3 removes the Cauchy constraint inherent in pair potentials, markedly improving elastic constants (C12 vs C44) and surface/vacancy energies, bringing performance close to MTP/GAP.
• Phonon spectra: UF2 reproduces DFT phonon dispersion significantly better than Morse and vastly better than LJ (which shows imaginary frequencies). Adding three-body terms (UF2,3) further closes the gap to MTP/GAP.
• Learning curves: With limited training data, simpler models (UF2) converge earlier but saturate at higher error; UF2,3 and qSNAP/SNAP require more data but reach lower errors, with UF2,3 comparable or superior to SNAP in energy accuracy.
• Melting temperature (two-phase method, W): DFT reference ≈ 3465 ± 105 K. Predictions: UF2 ≈ 3850 ± 68 K; UF2,3 ≈ 3651 ± 31 K; MTP ≈ 3961 ± 82 K; SNAP ≈ 3141 ± 54 K; qSNAP ≈ 3136 ± 63 K (calculation encountered instability at high T); Morse ≈ 2681 ± 45 K; EAM4 ≈ 4573 ± 78 K. UF2,3 is closest to DFT among tested models while remaining much faster than MTP/GAP/SNAP.

The UF approach directly addresses the central challenge of combining high accuracy with extreme speed and interpretability in atomistic simulations. By restricting the model to effective two- and three-body terms and using cubic B-spline bases with compact support, the method preserves physical interpretability (explicit V2 and V3 contributions) and enables analytical forces and efficient evaluation. Regularized linear least squares yields a convex, deterministic fit with smoothness control, improving robustness and avoiding overfitting. Results show that UF2 achieves accuracy comparable to sophisticated MLPs such as SNAP/qSNAP at the cost of simple empirical potentials, greatly extending feasible time and length scales. The three-body extension UF2,3 mitigates key limitations of pair potentials (e.g., violation of Cauchy relations), improving elastic constants and defect/surface energetics and approaching the accuracy of MTP and GAP at a fraction of their computational cost. The models also display reasonable extrapolative behavior in melting simulations despite the absence of liquid configurations in training, indicating robustness. Overall, the findings demonstrate that carefully designed low-order, spline-based linear potentials can redefine the speed–accuracy Pareto frontier for materials simulations.

The paper introduces ultra-fast, interpretable machine-learning potentials that combine spline-based effective two- and three-body interactions with regularized linear regression. For tungsten, UF2 matches the speed of the fastest empirical potentials while delivering accuracy comparable to SNAP/qSNAP; UF2,3 further approaches MTP/GAP accuracy with orders-of-magnitude lower computational cost. The models exactly recover known two-/three-body reference potentials and accurately predict derived properties such as phonons, elastic constants, surface and vacancy energies, and melting temperature. The approach yields fast, robust, and interpretable potentials suitable for large-scale, long-timescale simulations with modest training data. Future directions include active learning to enhance robustness and data efficiency, extending to four-body terms (for dihedrals relevant to molecular and biomolecular systems), and further optimization of the UF2,3 implementation. Open-source code and data facilitate adoption and benchmarking.

• Physical expressivity: Pair-only UF2 is constrained by central-force assumptions, enforcing Cauchy relations (e.g., C12 = C44 in cubic crystals), which can limit elastic property accuracy and defect/mechanical predictions; adding three-body terms alleviates but increases cost.
• Basis/derivative limitations: Natural cubic splines ensure continuous first and second derivatives but may be insufficient for properties requiring higher derivatives (e.g., thermal expansion, finite-temperature elastic constants).
• Data coverage: Training sets did not include liquid configurations; melting predictions test extrapolation and may be sensitive to dataset breadth. High-dimensional models (e.g., qSNAP) showed numerical instability at high temperatures, hinting at sensitivity to underrepresented configurations.
• Implementation and performance estimates: UF2,3 cost reported as bounds; current implementation not fully optimized.
• Generalizability: While tungsten benchmarks are strong, performance on more complex, multicomponent, or strongly directional systems may require higher body order (e.g., four-body) and more extensive data, potentially reducing the ultra-fast advantage.