FreeSeek
Talk to us
MAT · New Materials R&D

Agents that plan the next experiment from every past one

Materials teams sit on years of experiments locked in notebooks, spreadsheets and instrument folders. FreeSeek builds agents that structure this record, recommend formulations and process windows, and screen ideas against the patent landscape. Scientists keep control of what gets synthesized and what gets filed.

Systems we connect to
ELNLIMSMaterials ProjectPython/JupyterBoTorchpymatgenDerwent Innovation
Application scenarios Case studies
Industry context

Where materials R&D stands today

Most materials development still runs on expert intuition and one-factor-at-a-time trials. A new polymer blend, cathode coating or alloy grade can take hundreds of experiments over several years. The data from those experiments is often incomplete: process conditions live in personal notebooks, characterization results sit in vendor file formats, and failed runs are rarely recorded at all. Each new project restarts from partial memory, and knowledge leaves with the people who move on.

Public resources such as Materials Project, open literature and patent databases now hold millions of composition and property records, and Bayesian optimization has matured into usable open-source libraries. Few labs combine these with their own data. The gap is engineering work: cleaning historical records, linking samples to measurements, and wrapping models in a workflow that bench scientists trust. Agents can carry that integration work and keep it current as new results arrive.

Long development cycles

Qualifying a new material for batteries, aerospace or electronics often takes 5 to 10 years. Customers and regulators expect shorter timelines each product generation.

Fragmented experimental data

Recipes, process logs and test results are spread across ELN entries, Excel files and instrument PCs. Without a linked record, teams repeat experiments and cannot model what they already know.

Rising cost per experiment

Precursor prices, instrument time and skilled labor make every synthesis run expensive. Labs need each experiment to carry more information.

Dense patent landscape

Battery chemistries, engineering polymers and coatings are crowded with overlapping claims. A promising formulation can be blocked late, after significant scale-up spend.

Agent run · example

Watch an agent work a real task

Every step is a tool call against your own systems, logged and reviewable. The result goes to a named person for sign-off.

Agent run · examplerun_b40c98
TaskPropose the next 8 epoxy formulations to reach Tg above 150 °C with viscosity below 2 Pa·s.
  1. eln.search()
    Retrieved 312 past epoxy experiments with recipes, cure profiles and DSC results.
  2. bo.suggest()
    Fitted a surrogate model and ranked 8 candidates by expected improvement within cost limits.
  3. lims.draft_request()
    Drafted a synthesis and DSC test request for the 8 candidates.
Result · Batch plan sent to the project scientist for approval before synthesis.
Application scenarios

Four workflows, analysed step by step.

Each scenario shows how the work is done today, what the agent takes over, where a person signs off, and what we measure.

MAT-1Formulation and process-window recommendationMAT-2Literature and patent mining into composition–property datasetsMAT-3Characterization data interpretation and ELN structuringMAT-4Freedom-to-operate pre-screening
MAT-1

Formulation and process-window recommendation

The agent learns from historical experiments and proposes the next batch of formulations and process settings most likely to meet the target specification.

Integration depth Deep integration
Typical pilot8–12 weeks
How it works today

Scientists choose the next experiments from experience and a few spreadsheet trends. Designs of experiments are often full factorial or one variable at a time, which wastes runs in regions already known to fail. Historical data from earlier projects is rarely reused because units, sample names and test methods differ. Reaching a formulation that balances strength, viscosity and cost can take dozens of iterations and several months of bench time.

Systems used
ELNLIMSBoTorchPython/Jupyterlab automation scheduler
What the agent does
  1. Pulls past recipes, process parameters and measured properties from the ELN and LIMS, and harmonizes units, sample IDs and test methods.
  2. Fits a surrogate model with uncertainty estimates and checks it against held-out experiments before use.
  3. Proposes a ranked batch of next experiments with Bayesian optimization, honoring constraints such as raw-material lists, cost ceilings and equipment limits.
  4. Writes the batch as a design-of-experiments plan and, where an automated line exists, queues it through the lab scheduler.
  5. Ingests the new results, updates the model and reports progress toward the target and remaining uncertainty.
Human checkpoint
A project scientist approves every proposed batch before synthesis, checking safety, raw-material availability and chemical sense. The team lead decides when a candidate is good enough to move to pilot scale.
What we measure
  • experiments to reach target specification
  • share of proposed runs meeting spec
  • model prediction error on new runs
  • bench time per development cycle
MAT-2

Literature and patent mining into composition–property datasets

The agent reads papers and patents in a target domain and extracts compositions, processing conditions and measured properties into a structured, source-linked dataset.

Integration depth Moderate
Typical pilot6–8 weeks
How it works today

Researchers build literature datasets by hand, reading papers and copying values from tables and figures into spreadsheets. A review of a few hundred papers takes weeks, and coverage stays narrow. Units, test standards and sample preparation vary between sources, so the numbers are hard to compare. Patents are read separately by IP staff, and their examples seldom reach the technical team as usable data.

Systems used
Scopus APIEspacenetDerwent InnovationMaterials ProjectPython/Jupyter
What the agent does
  1. Searches literature and patent databases with a query agreed with the research team, and records the full search strategy.
  2. Extracts compositions, processing steps, test conditions and property values from text, tables and figure captions.
  3. Normalizes units and maps test methods to standards such as ASTM D638 or ISO 527, and flags values that are not comparable.
  4. Links every record to its source page and sentence, and assigns a confidence score.
  5. Publishes the dataset to the team workspace and merges it with internal data for modeling.
Human checkpoint
A domain scientist spot-checks a sample of extracted records against the original sources and signs off before the dataset is used for modeling or decisions.
What we measure
  • records extracted per week
  • extraction accuracy on audited samples
  • share of records with full process conditions
  • time to first usable dataset
MAT-3

Characterization data interpretation and ELN structuring

The agent reads raw XRD, SEM, DSC and spectroscopy files, drafts an interpretation, and files the results against the right sample in the ELN.

Integration depth Moderate
Typical pilot6–10 weeks
How it works today

Each instrument writes files in its own format to a local PC. Scientists copy files to shared drives, analyze them by hand in vendor software, and paste screenshots into notebook entries. Peak assignments, glass transition temperatures and particle sizes are typed in manually, often days later. Links between a sample, its recipe and its measurements break easily, which makes results hard to find, reuse or audit.

Systems used
ELNLIMSinstrument raw data filespymatgenICDD PDF-4+
What the agent does
  1. Watches instrument export folders and parses raw files and their metadata, including operator, method and calibration date.
  2. Matches each file to a sample ID and recipe in the ELN and LIMS, and queries the operator when the match is ambiguous.
  3. Runs standard analyses such as XRD phase matching, DSC transition detection and SEM particle-size statistics.
  4. Drafts an interpretation with key values, confidence and comparison to reference patterns or earlier batches.
  5. Writes a structured entry to the ELN with raw data, processing parameters and the analysis script version attached.
Human checkpoint
The analyst who ran the measurement reviews and accepts each interpretation before it becomes part of the official record. Unusual phases or anomalies are escalated to the characterization lead.
What we measure
  • time from measurement to ELN entry
  • share of measurements linked to a sample record
  • interpretation acceptance rate
  • analyst hours per sample
MAT-4

Freedom-to-operate pre-screening

The agent compares candidate formulations and processes against active patent claims and flags potential conflicts early, before scale-up money is spent.

Integration depth Moderate
Typical pilot6–8 weeks
How it works today

Formal freedom-to-operate opinions are expensive and usually commissioned only once a material is close to launch. Earlier in development, scientists rely on informal searches or none at all. When a conflict appears late, the team must redesign around it or license, after months of scale-up work. IP counsel receives requests with incomplete technical detail, which slows the review and raises its cost.

Systems used
Derwent InnovationEspacenetUSPTO Patent Public SearchCNIPA databaseELN
What the agent does
  1. Converts a candidate formulation or process into a structured description of components, ranges and steps.
  2. Searches active patents and published applications in the relevant jurisdictions, filtered by legal status and expiry date.
  3. Maps independent claims element by element against the candidate and highlights overlapping ranges.
  4. Produces a risk-ranked pre-screen report with claim charts and possible design-around directions.
  5. Routes medium and high-risk items to IP counsel with the full search record attached.
Human checkpoint
Patent counsel reviews every flagged item and makes all legal judgments. The pre-screen informs research priorities and does not replace a formal freedom-to-operate opinion.
What we measure
  • candidates screened before scale-up
  • late-stage patent conflicts
  • counsel hours per review
  • time from request to pre-screen report
Case studies

How engagements in this field have played out.

Illustrative cases: composites of typical deployments with client details removed. Figures show the range our pilots are designed to reach.

Bayesian optimization shortens a battery binder development cycle

Client: Battery materials producer, cathode and binder lines, 2 R&D centersRegion: East Asia
−41%formulation runs to reach target
−35%development cycle time
86%historical records recovered as usable data
Challenge

The team needed a water-based binder formulation for a high-nickel cathode that met adhesion, slurry viscosity and cycle-life targets. Six years of related experiments existed, but in three ELN systems and many spreadsheets with inconsistent units. Previous projects of this type took more than 18 months and around 300 formulation runs.

What we built
  • Harmonized about 2,400 historical experiments from the ELN, LIMS and spreadsheets into one linked dataset.
  • Built a multi-objective Bayesian optimization loop with cost and raw-material constraints.
  • Proposed weekly batches of 12 formulations, each approved by the project scientist before synthesis.
  • Fed coin-cell and adhesion test results back to the model automatically through the LIMS.
Timeline: 10-week pilot, production use in month 5

Structured characterization records for a coatings lab

Client: Industrial coatings manufacturer, 4 application labsRegion: Southern Europe
−62%time from measurement to ELN entry
−30%analyst hours per sample
97%measurements linked to a sample record
Challenge

Analysts spent much of their week moving DSC, FTIR and SEM files into reports by hand. Results were often filed days after measurement and sometimes against the wrong batch. During a customer audit, the lab struggled to show which method version and calibration produced specific results.

What we built
  • Connected the agent to instrument export folders and the existing ELN and LIMS.
  • Automated parsing, sample matching and standard analyses for DSC, FTIR and SEM.
  • Introduced analyst review and acceptance before any entry reached the official record.
Timeline: 8-week pilot, rollout to all 4 labs by month 6
Risk and compliance

IP, data integrity and export control in materials R&D

Trade secrets and IP protection

Formulations and process windows are often protected as trade secrets, so the agent runs with role-based access and full audit logs. Data is never used to train shared models, and outputs that may support a patent filing are routed to IP counsel first.

On-premises model deployment

For sensitive programs we deploy models inside the client network or a private cloud tenant. No experimental data leaves the client environment unless the client approves it.

Data provenance and reproducibility

Every recommendation records the data version, model version and analysis code that produced it, following FAIR data principles. Labs accredited to ISO/IEC 17025 can trace each reported result back to raw instrument files.

Export controls

Some materials and technologies fall under the Wassenaar Arrangement, the US EAR, EU Regulation 2021/821 or the China Export Control Law. Access to controlled project data is restricted by user nationality and location rules that the client defines with its trade compliance team.

Questions we hear
How much historical data do we need before optimization is useful?

A few dozen well-documented experiments in the target system are often enough to start, because Bayesian optimization is designed for small datasets. The agent also reports model uncertainty, so the team can see when more exploratory runs are needed.

Does the agent run experiments without human approval?

No. The agent proposes batches and drafts lab requests, and a scientist approves each batch before synthesis. In automated labs the scheduler only receives approved plans.

Can the system work with our existing ELN and instruments?

Yes. We connect to common ELN and LIMS platforms through their APIs and read instrument files from export folders. Where no API exists, we build a file-based connector during the pilot.

Does the patent pre-screen replace a freedom-to-operate opinion?

No. It helps the research team spot likely conflicts early and gives counsel a clear technical starting point. Legal conclusions remain with qualified patent attorneys.

Have a workflow like these?

Send us a short description of how the work is done today. We reply within two business days.

Start a conversation
Other industries
MFGAdvanced ManufacturingENGEngineering DesignMEDHealthcare ServicesFINFintechMKTMarketing