Materials science workflows and evaluation · 2026
Materials science workflows, tools & evaluations.
Compare published evidence for materials research workflows, from simulations and experimental analysis to agent benchmarks.
LEADERBOARD · NOVEMBER 2026
Materials science workflow and evaluation providers
7 ranked specialists · 7 tracked (NR). Expand an entry for its role, evidence and sources. Equal assessments share a rank.
01 Matlantis↗Supplies atomistic simulation for materials screening and experimental follow-up. SimulationExperimental analysis Methanol catalyst screening and verificationPublished evidence CommercialTokyo, Japan Corroborated
Supplies atomistic simulation for materials screening and experimental follow-up.
The published workflow compares catalyst models, screens dopants and tests selected candidates experimentally. A Seoul National University team separately describes its use of Matlantis in catalyst research. These results concern simulation and materials development.
02 Atinary↗Supplies active-learning software for closed-loop catalyst experiments. Experimental analysis SwissCat+ catalyst study and hydroformylation optimizationPublished evidence CommercialEpalinges, Switzerland Corroborated
Supplies active-learning software for closed-loop catalyst experiments.
ETH Zurich and Atinary documented a catalyst workflow. A separate dsm-firmenich collaboration used Atinary’s platform to choose conditions and run 88 hydroformylation experiments. The outcomes concern experimental optimization.
=3 Citrine Informatics↗Provides materials informatics tools for selecting and testing experimental candidates. Experimental analysis Closed-loop nanoparticle synthesis and formulation studiesPublished evidence CommercialRedwood City, US Documented
Provides materials informatics tools for selecting and testing experimental candidates.
Citrine documents candidate selection followed by experimental feedback in nanoparticle synthesis and a separate semiconductor formulation comparison. The studies describe materials development workflows. Agent training gains were not verified.
=3 Collinear AI↗Evaluates agents on computational and experimental materials research workflows. BenchmarksSimulationExperimental analysis MatFlowBenchPublished evidence CommercialSan Francisco, US Documented
Evaluates agents on computational and experimental materials research workflows.
MatFlowBench evaluates 45 private workflows across computational and experimental research. Agents work within fixed scientific and resource constraints. Task-specific verifiers assess submitted results. Experimental tasks use measurement archives. Outside execution and controlled training gains were not verified.
=3 Radical AI↗Builds materials extraction benchmarks and tests candidate materials in an autonomous lab. BenchmarksExperimental analysisLiterature extraction LitXAlloy and semiconductor validationPublished evidence CommercialNew York, US Documented
Builds materials extraction benchmarks and tests candidate materials in an autonomous lab.
LitXAlloy evaluates extraction of experimental records from alloy papers using structured checks. In a separate workflow, Radical reports making six model-proposed materials and measuring the target electronic property. These are distinct contributions. No outside run of LitXAlloy was verified.
06 MaterialsZone↗Supplies materials data infrastructure and recommends experiments from recorded results. Experimental analysis Active-learning optimization and spectroscopy data workflowsPublished evidence CommercialTel Aviv, Israel Documented
Supplies materials data infrastructure and recommends experiments from recorded results.
MaterialsZone describes an iterative experiment-selection workflow and reports a reduction in experimental iterations in a use case. A separate consortium case documents spectroscopy data handling across institutions. The reported savings are case-specific, and independent replication was not verified.
07 Mat3ra↗Provides reproducible workflows for building and screening materials structures. Simulation Mat3ra-2D methods and worked examplesPublished evidence CommercialWalnut Creek, US Documented
Provides reproducible workflows for building and screening materials structures.
Mat3ra-2D publishes a framework for constructing realistic 2D materials and interfaces, with documented treatment of disorder and defects. This is a structure-design workflow rather than a measured agent-training result.
NR Ansys Granta↗Supplies materials informatics tools for property prediction and design validation. SimulationExperimental analysis Granta MI model-based materials developmentTracked incumbent CommercialCanonsburg, US Documented
Supplies materials informatics tools for property prediction and design validation.
The published case describes using standardized materials test data for property prediction and reducing prototype materials and tests. It documents a materials informatics workflow rather than an agent evaluation.
NR Edison Scientific↗Builds agents for scientific literature search and data analysis. Literature extractionBenchmarks Platform agent evaluationsEvidence gap CommercialSan Francisco, US Limited
Builds agents for scientific literature search and data analysis.
Published platform evaluations include bioinformatics and scientific analysis. The sources reviewed did not establish a completed materials-specific workflow with comparable methods and outcomes.
NR Kebotix↗Offers materials design, property prediction and instrument orchestration. SimulationExperimental analysis ChemOS and automated experiment case studyEvidence gap CommercialCambridge, US Limited
Offers materials design, property prediction and instrument orchestration.
Kebotix documents experiment selection and remote submission to NCATS high-throughput equipment. The reviewed case concerns chemistry optimization. A completed, attributable materials-specific result with comparable verification was not established.
NR Microsoft Research↗Develops generative models for inorganic materials design under property constraints. SimulationExperimental analysis MatterGen property-guided materials designTracked incumbent Commercial research labRedmond, US Documented
Develops generative models for inorganic materials design under property constraints.
MatterGen reports generating crystal structures under target property constraints and describes an experimental synthesis example. The released implementation supports reproduction of model workflows. Its results measure materials generation and do not establish general research-agent performance.
NR Orbital Industries↗Develops interatomic models for atomistic materials simulation. Simulation Orb simulation and crystal stability evaluationsEvidence gap Commercial + OSSLondon, UK Limited
Develops interatomic models for atomistic materials simulation.
Orb has published model-level simulation results. This review did not establish a comparable supplied agent task suite or a verified end-to-end materials research workflow for this snapshot.
NR Phaseshift Technologies↗Uses simulation and experimental data to design materials for production processes. SimulationExperimental analysis Commercial alloy design workflowEvidence gap CommercialToronto, Canada Limited
Uses simulation and experimental data to design materials for production processes.
The company describes simulation-guided selection and commercial equipment qualification. The reviewed source did not provide a named completed study with inspectable methods and measured outcomes.
NR Schrödinger↗Provides physics-based simulations and active learning for materials design. SimulationExperimental analysis OLED active learning and LiveDesign validationTracked incumbent CommercialNew York, US Documented
Provides physics-based simulations and active learning for materials design.
Schrödinger documents an OLED candidate selection workflow using molecular simulation and active learning. Its LiveDesign case compares calculated orbital energies with experimental measurements. The studies evaluate materials design methods and do not establish agent training gains.
Methodology
The index covers commercial specialists that supply materials research workflows, simulation tools, experimental systems or agent evaluations. A rank requires a completed technical contribution with verifiable methods, outcomes and attribution. This snapshot covers the providers reviewed here.
We assess the strongest documented contribution for each factor. Equal overall assessments share a rank. Broad incumbents are tracked separately. Model-only results and adjacent scientific work are tracked where relevant.
Open, restricted and private work face the same evidence standard. Company size, customer logos and announcements carry no weight.
Documented means the evidence meets the threshold. Corroborated adds confirmed outside execution or use. Limited indicates an unresolved evidence gap.
- 01
Technical substance, depth and scope
Credit completed work across distinct scientific outcomes, such as property estimation, candidate selection or characterization. Several calculations toward one result count as one workflow.
- 02
Scientific realism
Look for realistic materials, software, measurement artifacts and research constraints. Examples include sample purity, instrument limits and processing conditions that affect the result.
- 03
Verification rigor
Check reference values, units, numerical tolerances and failure cases. Credit convergence tests, independent measurements and checks that distinguish plausible output from a supported conclusion.
- 04
Training utility
Distinguish static examples, repeatable evaluations, runnable training tasks and controlled gains on held-out materials or workflows. Standalone simulation accuracy measures a different capability.
- 05
External validation
Distinguish provider-run work, confirmed collaboration, completed outside use and independent replication. Multiple references to one run count once.
Materials science AI FAQ
How to compare materials workflows, agent evaluations and experimental evidence.
What does a materials workflow benchmark measure?+
It measures whether an agent can carry a bounded materials research question through to a result supported by evidence. Some tasks center on simulation, such as choosing a method and interpreting its output. Others center on published or recorded data, such as extracting measured values from papers. Check whether the benchmark uses recorded data or physical lab execution, and what the grader verifies.
How does a workflow benchmark differ from a materials property prediction benchmark?+
A prediction benchmark gives a model a defined input and scores a property estimate. A workflow benchmark can require the agent to choose a calculation, diagnose failure, interpret output and report an answer under time or compute constraints. The results answer different questions about a system’s capabilities.
What makes a computational materials task realistic?+
A realistic task specifies the material, target property, permitted tools and resource budget. The agent may need to select convergence settings, check whether the output represents the intended physical state and explain its uncertainty. A calculation that exits successfully can still produce the wrong scientific result.
How should an agent be evaluated on experimental materials work?+
For a task built on recorded data, check whether the agent selects and interprets the right measurements, reconciles units and reports limitations. For physical lab execution, record instrument settings, sample identity, safety limits and measured outcomes. Report which of these settings the evaluation actually exercised.
What makes a materials workflow reliably verifiable?+
A verifier should check the requested property or conclusion against an appropriate reference and tolerance. It can also inspect whether the run used the intended material, settings and units. Include failed or ambiguous examples to see whether the verifier rejects an unsupported answer.
How can teams avoid leakage in materials science evaluations?+
Separate training and evaluation by material family, study or workflow where practical. Keep reference results away from the agent unless retrieval is part of the task. Audit near-duplicates and record which databases, papers and artifacts the agent could access.
What shows that a workflow improves agent training?+
Look for a comparison with a stated baseline under the same evaluation conditions, using tasks held out from training. Gains on unseen material families or workflow types provide stronger evidence of transfer. Distinguish improved agent behavior from a change to the underlying simulation model.
How should a lab compare materials workflow providers?+
Start with one computational or experimental workflow the team actually runs. Ask for the inputs, execution record, verification criteria and failure examples. Check tool access, reproducibility, licenses and whether the provider can support the intended use, such as simulation, model evaluation or post-training.