5,008 DFT-grade property predictions. 13.5 seconds. The science that makes it possible.

The challenge

Materials discovery teams in industry and academia face a structural bottleneck that has nothing to do with ideas and everything to do with the realistic alternatives for getting accurate property predictions. None of the four standard pathways scale gracefully with the questions a modern discovery program needs to ask.

• High-throughput DFT on an HPC cluster routinely consumes 50,000 to 500,000 CPU-hours for a 100–1,000-material screen across an electronic, vibrational, and elastic property suite. Wall-clock measured in weeks to months. Annualized capability cost (cluster + computational personnel): typically several hundred thousand euros per group.
• In-house ML pipelines require thousands of labeled examples per property, an ML team to maintain them, drift retraining cycles two to four times a year, and they degrade sharply on chemistry far from the training distribution — precisely the regime that matters for discovery.
• Stitched commercial stacks (Schrödinger Materials Science Suite + database-lookup tools + internal ML for the gaps) carry six-figure annual licenses, integration engineering overhead, and fall back to full DFT for any composition not already deposited in the underlying databases.
• Database lookup in AFLOW, JARVIS, or the Materials Project is fast and free — but only if the compound of interest is already there. For novel chemistry, the DFT calculation has to be commissioned anyway.

The practical effect is that programs constrain the scope of their screening campaigns to fit available compute capacity and integration overhead. Many lines of inquiry remain unattempted because the cost structure of the underlying tools rules them out.

Study design

The study reproduces a generic 16-property screen across the FluxMateria materials reference cohort — the same 313 materials and 16 properties used in the publicly audited materials-universal benchmark. Each material is evaluated on the full property suite. Wall-clock time is measured end-to-end, including I/O. The run is deterministic: identical inputs produce bit-identical outputs across machines.

Results overview

FluxMateria completed all 5,008 property predictions in 13.5 seconds of wall-clock time, with accuracy matching the publicly audited materials-universal benchmark: 1.17% MAPE on family holdout. On the same hard split, the named alternatives report 36.07% (AFLOW), 10.92% (JARVIS), and 18.42% (MatBench) — FluxMateria is between 9 and 31 times more accurate than each of them.

Wall-clock measured on a single CPU core for deterministic reproducibility. Production deployment scales horizontally. Accuracy figure from publicly audited materials-universal benchmark, sealed to commit f4fb848.

Total cost of ownership: the annualized view

Enterprise discovery programs do not buy a single screen — they buy a capability. The relevant comparison is the annualized cost of running materials property prediction as an ongoing function: licenses, personnel, compute, integration, and the engineering cycles spent keeping the pipeline alive. Below, the four realistic alternatives a materials program faces today, costed honestly.

Annualized cost ranges represent typical industry benchmarks for ongoing capability: HPC pathway includes cluster allocation plus 1–2 computational personnel; ML pathway includes a small ML team plus data labeling plus drift retraining; stitched commercial stack includes typical pharma/materials site licenses ($200K–$500K range) plus integration engineering plus residual cluster compute for novel compositions. These are not headline list prices; they reflect what enterprise programs actually spend over a fiscal year. FluxMateria pricing is enterprise-tiered and disclosed under NDA.

Accuracy: validation against the public benchmark

The accuracy of the engine is validated against the publicly audited materials-universal benchmark — same engine, same property suite, same cohort, frozen to commit f4fb848. The numbers below come directly from the frozen JSON manifest released on that page.

Source: materials_physics_external_family_holdout_2026-02-24.json, commit f4fb848. NA on adapter rows = the named adapter does not score those tests.

In short: 9 to 31 times more accurate than the named alternatives on the same hard split. Not a small accuracy concession in exchange for speed — an accuracy gain on top of the speed gain.

Operational implications

First-principles accuracy at workflow latencies enables a class of design and review operations that conventional pathways constrain by compute budget or output-schema mismatch.

Honest scope

DFT-grade accuracy at workflow speed is a strong claim. It deserves a clean fence around what is and is not in scope.

The accuracy and cost numbers in this case study apply to the cohort and property suite documented in the materials-universal benchmark. Extending to new families is a documented engineering process — not a free claim.

Conclusion

FluxMateria delivers DFT-grade accuracy across a 16-property materials suite at workflow latencies, with no fitting and no training data. On the publicly audited family-holdout split, overall MAPE is 1.17% — between 9 and 31 times closer to experiment than the AFLOW, JARVIS, and MatBench adapters scored on the same materials. Annualized capability cost sits below the four enterprise pathways a materials program faces today: high-throughput DFT on cluster, in-house ML pipelines, stitched commercial stacks, and database lookup with DFT fall-back.

Accuracy and throughput are co-derived outputs of a single first-principles physics model. The model does not require functional, basis-set, or k-point selection; both metrics are consequences of the same axiomatic derivation, validated against external public datasets and frozen to a downloadable JSON manifest.

Technical specifications