solver.press

Untested hypothesis

Integrating the O(N) sparse differentiation Hessian algorithm with long-range machine learning interatomic potentials (such as MACE-POLAR or LOREM) enables the calculation of exact analytic Hessians for defect structures in \text{SrTiO}_3 supercells containing over 10,000 atoms in less wall-clock time than standard numerical differentiation using a short-range baseline MACE model.

PhysicsOct 5, 2026Evaluation Score: 66%

Adversarial Debate Score

57% survival rate under critique

Expert panel critique

Independent views, each critiquing the hypothesis on its own — the score rewards genuine disagreement and discounts consensus.

Gemini: Strengths: The hypothesis is highly plausible and falsifiable; integrating O(N) sparse automatic differentiation with long-range machine learning interatomic potentials (like MACE-POLAR) directly bypasses the O(N^2) scaling of standard numerical differentiation for large-scale (10,000...
Mistral: The hypothesis is well-grounded in validated O(N) sparse differentiation methods and long-range MLIPs, with strong mechanistic plausibility and falsifiability. However, it lacks direct empirical validation in the owner’s experiments (which focus on unrelated precision/BO topics) and requires expl...
Claude: The hypothesis is falsifiable (a wall-clock benchmark), but it rests on a fundamental conflict: the cited sparse-differentiation papers explicitly depend on locality of the MLIP, giving a sparse Hessian pattern that colouring exploits to reach O(N) cost, while long-range models like MACE-POLA...
ChatGPT: The claim is falsifiable and plausibly motivated, but the cited \(O(N)\) exact-Hessian method appears predicated on locality/sparsity, whereas long-range interactions can produce dense Hessian couplings and undermine that scaling. No relevant validated owner experiment or direct \(>10{,}000\)-ato...

Related patents (prior art)

This hypothesis overlaps subject matter covered by existing third-party patents. It is published as research, not as a patentable claim of ours.

Supporting Research Papers

Formal Verification

Z3 logical consistency:✅ Consistent

Z3 checks whether the hypothesis is internally consistent, not whether it is empirically true.

Experimental Validation Package

This discovery has a Claude-generated validation package with a full experimental design.

Precise Hypothesis

For defect-containing SrTiO3 supercells with N > 10,000 atoms, computing the exact analytic Hessian via sparse automatic differentiation (O(N) sparse-forward/reverse-mode AD exploiting the interatomic potential's local connectivity graph) applied to a long-range-aware machine-learned interatomic potential (MLIP) such as MACE-POLAR or LOREM will complete in less wall-clock time (measured on identical hardware, single-node, same parallelization budget) than finite-difference numerical Hessian estimation (central difference, displacement δ=0.01 Å, 6N single-point evaluations) using a short-range-cutoff baseline MACE model of comparable accuracy (validation force RMSE within ±15%), while achieving Hessian element-wise agreement with a DFT-reference Hessian (PBEsol, tight settings) of Pearson r ≥ 0.95 on a held-out 200-atom defect cluster subsystem.

Disproof criteria:
  1. Wall-clock crossover does not occur by N=10,000 (sparse exact Hessian is still slower than numerical short-range baseline at this size) — falsifies the core speed claim.
  2. Sparse AD Hessian fails accuracy threshold (Pearson r < 0.95 vs DFT reference on the 200-atom subsystem) — falsifies the "exact" utility claim even if fast.
  3. Sparsity pattern assumption breaks down: measured nonzero Hessian block density exceeds 10% at N=10,000, causing sparse AD runtime to approach or exceed dense AD runtime (O(N) claim falsified).
  4. MACE-POLAR/LOREM energy surface is not twice-differentiable in practice (NaN/Inf in second derivatives, or AD framework cannot backprop through the long-range solver twice) — falsifies feasibility, not just performance.
  5. Memory requirements exceed available hardware (>80GB) before N=10,000 is reached, making the comparison moot on standard single-GPU infrastructure.

Spine & Adversarial ReadReady for validation

“This hypothesis tests whether combining an O(N) sparse automatic-differentiation Hessian algorithm with a long-range-aware machine-learned interatomic potential computes exact Hessians for >10,000-atom SrTiO3 defect supercells faster than numerical differentiation with a short-range MACE baseline.”

  • highThe comparison may be structurally unfair: numerical differentiation cost scales with the choice of baseline model's inference cost, not with the Hessian algorithm itself — a short-range MACE forward pass is cheaper per-call than a long-range MLIP forward pass, so the 'speedup' could simply reflect per-call cost differences (O(N) short-range inference vs O(N) or worse long-range inference) rather than the sparse-AD-vs-numerical-differentiation algorithmic advantage being claimed.
    Partially addressed by step 9 (sensitivity analysis) but not fully resolved in this EVP. A cleaner ablation would additionally compare sparse-AD-Hessian vs numerical-Hessian using the SAME long-range model, isolating the algorithmic (AD vs finite-difference) contribution from the model-choice (long-range vs short-range) contribution. This two-factor design should be added before claiming the result supports the stated hypothesis as currently phrased, which conflates both factors.
  • mediumWhy SrTiO3 specifically, and why PBEsol/VASP as DFT reference, and why MACE-POLAR/LOREM specifically rather than other long-range MLIPs (e.g., 4G-HDNNP, Allegro with long-range extensions, or NequIP variants)? The methodology does not justify these specific choices over alternatives, which is a common point of rejection.
    SrTiO3 is a reasonable choice as a canonical polar perovskite with well-characterized defect chemistry and existing MLIP training data, but this justification is not stated in the original hypothesis and should be made explicit. PBEsol is standard for perovskite lattice properties but a harder DFT reference (e.g., hybrid functional) might be needed if PBEsol itself has known errors for SrTiO3 defect states — this is an unresolved gap. The choice of MACE-POLAR/LOREM over alternative long-range MLIPs is apparently driven by availability/familiarity rather than a documented comparative justification; this EVP recommends an explicit model-selection rationale be added as a pre-registration step before full-scale execution.
  • mediumExtrapolating or even directly measuring a single crossover point at N=10,000 for one defect type (oxygen vacancy) and one material (SrTiO3) provides very limited generalizability evidence for the broader claim implied by the impact statement (accelerating 'computational materials discovery for perovskites and other polar materials' generally).
    Acknowledged and not resolved within this minimum-viable protocol's budget. The MVT is explicitly scoped to one material/one defect type for tractability; generalization to other perovskites (BaTiO3, PbTiO3) and defect types (antisites, polarons) would require a follow-on validation phase, which should be scoped separately once the core N-scaling claim is confirmed or refuted.

Experimental Protocol

Minimum viable test (MVT): A 4-point scaling study at N ∈ {500, 2,000, 5,000, 10,000} atoms using one SrTiO3 oxygen-vacancy defect supercell family, comparing (a) sparse exact Hessian via MACE-POLAR (or LOREM if available pretrained) against (b) numerical central-difference Hessian via short-range MACE-MP-0 (or equivalent published short-range baseline), measured on identical single-GPU hardware (A100-80GB), with DFT Hessian validation only at N=500 (tractable) and N≈1,000 subsystem extraction at N=10,000 for accuracy cross-check.

Required datasets:
  • SrTiO3 bulk + defect structures: generated in-house (VASP/Quantum ESPRESSO relaxed geometries) or sourced from Materials Project (mp-5229 SrTiO3 entries) and OQMD for base structures; defects introduced via supercell construction (ASE).
  • Pretrained MLIPs: MACE-MP-0 (short-range baseline, publicly released weights) and MACE-POLAR or LOREM (long-range variant — availability must be confirmed; if not publicly released, requires training from scratch, which materially changes cost/timeline — flagged as a critical dependency risk).
  • DFT reference Hessians: computed via VASP or Quantum ESPRESSO with PBEsol functional, tight convergence (10⁻8 eV force convergence), phonon/DFPT or finite-difference Hessian on the 200–1,000 atom subsystems only (full 10,000-atom DFT Hessian is computationally infeasible and is NOT required — only subsystem cross-validation).
  • Sparse AD framework: JAX with experimental sparse Jacobian support, or a custom implementation using graph-colouring-based sparse Hessian algorithms (e.g., building on jax.experimental.sparse, or PyTorch's functorch/torch.func with sparsity-aware vmap batching).
  • Compute environment: single-node multi-GPU cluster (A100-80GB ×1–4) for MLIP inference + AD; CPU cluster (64–128 cores) for DFT reference calculations.
Success:
  • Sparse exact Hessian wall-clock time at N=10,000 is ≤ 80% of numerical short-range baseline wall-clock time (meaningful, not marginal, speedup).
  • Hessian accuracy: Pearson r ≥ 0.95, RMSE ≤ 15% of mean |Hessian element| vs. DFT reference on subsystem validation.
  • Phonon frequencies (lowest 20 modes) from sparse-exact Hessian within 5% of DFT-derived frequencies on the validated subsystem.
  • Measured sparsity density < 2% nonzero blocks at N=10,000, confirming sub-quadratic scaling empirically (fit exponent ≤ 1.3 in log-log time-vs-N plot).
  • Crossover point (if any) occurs at N ≤ 10,000, i.e., directly observed rather than extrapolated.
Failure:
  • No crossover observed by N=10,000 and scaling fits suggest crossover (if any) exceeds N=50,000 (impractical regime).
  • Accuracy falls below r=0.85 vs DFT reference — indicates the "exact" claim is not delivering usable physical fidelity.
  • Long-range MLIP (MACE-POLAR/LOREM) is unavailable/untrainable within budget, forcing indefinite postponement — treated as inconclusive failure requiring re-scoping.
  • AD framework cannot differentiate twice through the long-range solver (numerical instability, NaNs) in >20% of test configurations.
  • Memory footprint exceeds 80GB before N=10,000, preventing single-GPU validation entirely.

1,850

GPU hours

75d

Time to result

$42,000

Min cost

$165,000

Full cost

ROI Projection

Commercial:

Direct applicability to photovoltaic perovskite stability screening (defect-driven degradation pathways), ferroelectric memory device design (polar instability prediction), and solid-state electrolyte defect chemistry. Potential commercial licensees: materials-discovery platform companies (e.g., computational screening SaaS), battery/semiconductor manufacturers needing defect stability prediction, and national lab/HPC software ecosystems (VASP, Quantum ESPRESSO plugin ecosystem). Method is also generalizable beyond SrTiO3 to any polar/long-range-coupled crystal system, broadening addressable market to all MLIP-based high-throughput screening pipelines (estimated TAM for computational materials screening tools: $200-500M research-tooling market).

🔓 If proven, this unlocks

Proving this hypothesis is a prerequisite for the following downstream discoveries and applications:

  • 1high-throughput-defect-screening-perovskites
  • 2mace-polar-production-release-validation
  • 3ferroelectric-phonon-instability-screening-pipeline

Implementation Sketch

# Pseudocode: Sparse Exact Hessian Benchmark Pipeline

for N in [500, 2000, 5000, 10000]:
    structure = build_SrTiO3_defect_supercell(N, defect_type="O_vacancy")
    structure = relax(structure, calculator=MACE_shortrange)

    # --- Sparse exact Hessian path ---
    model_lr = load_pretrained("MACE-POLAR")  # or LOREM
    graph = build_interaction_graph(structure, cutoff_sr=model_lr.cutoff,
                                     longrange_module=model_lr.ewald_head)
    sparsity_pattern = graph_coloring_hessian_pattern(graph)  # O(N) colors expected

    t0 = time()
    H_sparse = sparse_hessian_AD(model_lr, structure, pattern=sparsity_pattern)
    t_sparse = time() - t0
    mem_sparse = peak_memory()

    # --- Numerical baseline path ---
    model_sr = load_pretrained("MACE-MP-0")
    t0 = time()
    H_numerical = finite_difference_hessian(model_sr, structure, delta=0.01)
    t_numerical = time() - t0
    mem_numerical = peak_memory()

    log_result(N, t_sparse, t_numerical, mem_sparse, mem_numerical,
               sparsity_density(H_sparse))

# --- Accuracy validation (subsystem only) ---
cluster = extract_embedded_cluster(structure_N10000, radius=8A)  # ~200-1000 atoms
H_dft = compute_dft_hessian(cluster, functional="PBEsol", code="VASP")
H_sparse_cluster = sparse_hessian_AD(model_lr, cluster)
pearson_r, rmse = compare_hessians(H_dft, H_sparse_cluster)
phonon_compare = compare_phonon_modes(H_dft, H_sparse_cluster, n_modes=20)

report(scaling_curves, crossover_point, pearson_r, rmse, phonon_compare)
Abort checkpoints:
  1. Day 10: Confirm MACE-POLAR/LOREM availability and basic forward-pass functionality on SrTiO3. If unavailable and training is infeasible within budget, abort/re-scope to a different long-range MLIP.
  2. Day 25: After N=500 and N=2,000 runs, check preliminary scaling exponent. If sparse method scaling exponent > 1.5 (not meaningfully sub-quadratic), abort before investing in N=10,000 compute.
  3. Day 35: After DFT reference Hessian computed on N=500 subsystem, check accuracy (Pearson r). If r < 0.85, abort — "exact" claim is not delivering physical fidelity regardless of speed.
  4. Day 55: After N=5,000 run, check memory trajectory. If extrapolated memory at N=10,000 exceeds 80GB by >2x, abort or pivot to multi-GPU sharding (budget re-negotiation required).

Source

AegisMind Research
Need AI to work rigorously on your problems? AegisMind uses the same multi-model engine for personal and professional use. Get started