The AI research environment
for materials science.
Domain-specialized agents that compress the materials R&D loop from months toward days — from literature and discovery through to synthesis.
Building toward the self-driving lab for materials R&D.
01 — THE PROBLEM
Most of the slowness isn't the science. It's the loop around the science.
A materials R&D team spends enormous time searching fragmented literature and databases, designing experiments, optimizing synthesis and process parameters, and turning raw instrument output into defensible decisions. Each full pass — hypothesis to validated result — takes months. That bottleneck compounds across a whole lab, a whole company, a whole decade of needed breakthroughs.
Months
per hypothesis-to-result pass
6+
characterization techniques to reconcile by hand
Fragmented
tools, databases, file formats, no shared context
02 — WHY GENERAL AI DOESN'T FIX IT
A researcher can't hand ChatGPT their actual work.
General tools don't speak the instrument file formats — .rasx, .mpr, .vms, BELMaster CSV.
They don't know the characterization conventions, and can't run the analysis on real data.
They can't reach the right scientific databases — crystal structures, diffraction references, the literature.
They never hold the project — context is lost the moment the session ends.
Materials science needs domain-specialized agents with the right tools, data access, and project memory — not a general chatbot.
03 — WHAT WE'RE BUILDING
An Integrated Materials Environment.
Create a project, bring your data, and work alongside an AI agent specialized for materials science. It reads and writes the project's files, runs real analysis in a sandbox, searches the literature, consults a library of materials-science skills, and keeps the whole project in context across sessions. It is not a chatbot with a science skin — it is the place the research lives.
FILES
- ▸ characterization/
- xrd/ Pt-MoS₂.rasx
- xps/ Pt-MoS₂.vms
- raman/ spectrum.txt
- ec/ HER_LSV.mpr
- ▸ literature/
- ▸ figures/
SESSION · Analyze XRD pattern
Generated 2 files in this session
Illustrative — workspace shown is a design preview.
04 — THE RESEARCH LOOP
Ten nodes. We earn each one by shipping the last.
The end state is the full autonomous research loop. We've mapped it as ten nodes — four work today, on real data. The platform fills in the rest, node by node.
- 01Literature & prior art
- 02Hypothesis generation
- 03Experiment & control design
- 04Synthesis recipe & protocol
- 05Sample fabrication
- 06Characterization
- 07Data analysis & interpretation
- 08Cross-technique validation
- 09Manuscript & figures
- 10Research direction update
Hover a node on the ring to trace the loop.
05 — THE HONEST THREE TIERS
We claim a real tool today — and a credible path to the lab.
WORKING NOW
Runs today, on real data
- — Characterization analysis (XRD, XPS, Raman, BET, TEM, EC)
- — Literature triage and synthesis
- — Cross-technique sanity-checking
- — Technical reporting and review
BUILDING NEXT
On the near roadmap
- — Hypothesis generation
- — Experiment and control design
- — Synthesis recipe and protocol tooling
THE HORIZON
The autonomous lab
- — Reads literature, forms hypotheses
- — Plans and coordinates experiments
- — Interprets results, updates its own direction
We are not claiming the autonomous lab today — we are claiming a real, used tool today and a mapped path to the lab.
06 — WHAT MAKES STOICH DIFFERENT
Four things a general tool can't do.
01
MLIP-first computation
Machine-learned interatomic potentials now hit near-DFT accuracy without HPC. Property prediction in minutes, DFT as the verification layer — materials' “AlphaFold moment.”
02
Instrument-format fluency
The agent natively reads .rasx, .mpr, .vms, BELMaster CSV and more. Generic AI tools simply cannot open these files.
03
The create-skill flywheel
Power users package their lab's workflow as a reusable skill. Instrument communities are small and tight — one lab's Rigaku skill is useful to twenty groups on the same instrument.
04
Cross-paper benchmarking
The agent re-derives a paper's overpotential from raw data and benchmarks it against twenty others. A verifiable demo beats an impressive one.
07 — WHO IT'S FOR
Built for the teams that move materials from idea to product.
Stoichtargets the slow, expensive parts of materials R&D — starting with synthesis and process optimization, where every wasted iteration burns weeks of lab time. The same agents serve a corporate R&D group, a national lab, or a hard-tech startup.
Energy storage & batteries
Electrode and electrolyte formulation, cycle-life screening, faster cell iteration.
Catalysis & clean energy
Catalyst discovery, activity benchmarking, and synthesis tuning straight from raw data.
Semiconductors & electronics
Thin films, 2D materials, and device-grade characterization at scale.
Structural & advanced materials
Alloys, composites, and process-property optimization across large parameter spaces.
Every project compounds into a provenance-rich record of how a material was actually made — the dataset that trains the autonomous lab.
08 — LANDSCAPE
Where Stoich sits.
Phylo / Biomni Lab
The same thesis for biology — A16Z and Menlo backed. Validates the shape. We are the materials counterpart.
Autonomous synthesis labs
A-Lab, Periodic, Lila — robotic, capital-heavy. A different end of the bench. We start software-first.
Materials Project
Computational databases and tooling we build on — not compete with.
The India angle
An India-based team building deep-tech R&D infrastructure, close to a fast-growing materials and manufacturing base.
09 — ROADMAP
Coming soon.
Hypothesis Generator
SOONQuestion to ranked hypotheses + control-matrix design
Recipe / Protocol Studio
SOONSynthesis recipe to scaled batch protocol + sanity check
Sub-agent Dispatch
SOONSpawn agents to work tasks in parallel inside a project
MCP Server Pack
SOONElectronic lab notebook, procurement, instrument booking
Resources Browser
SOONCrystal-structure, diffraction & literature databases
Skill Hub
SOONBrowse, enable and share community materials skills
Cloud-HPC DFT
SOONQueued DFT verification beyond MLIP screening
Robotic Instrument Bridge
SOONAPI to physical XRD, XPS and UV-Vis instruments
Memory Across Projects
SOONThe agent learns a researcher's recurring patterns
Watch the research loop run on real data.
Open the demo project — real multi-technique characterization data for a Pt-MoS₂ catalyst. Then bring your own.