An AI research agent for materials

MolAgent
Accelerating materials innovation with AI.

MolAgent is an AI research agent for materials science. It brings the newest scientific methods into working use, conducts complete simulation studies, and returns results you can trust.

Request Early Access → See How It Works

An expert across every level of theory

methods.proposed · curated_pipelines

Quantum

DFT · ab initio · TD-DFT

→

Atomistic

classical MD · force fields

→

Coarse-Grained

MARTINI · DPD · mesoscale

→

ML-Accelerated

MACE · NequIP · Allegro

→

The Agent

proposes · runs · analyzes

Every scale has its own codes, its own formats and its own traps — and a real project combines several of them. MolAgent is fluent across every level of theory. It doesn't decide for you: it proposes strategies, explains the trade-offs, and once you have chosen, it sets up the calculation, launches it, monitors progress and interprets the output.

describe · launch · arbitrate

4 levels

Theory levels covered

12 mo → 1 wk

Expert study, reproduced

100% on-prem

Runs on your hardware

0 outbound

Your data never leaves

01 / the bottleneck

Simulation is powerful and expensive to run properly.

The cost is not the computing. It is the expert knowledge required at three separate stages: setting the calculation up, executing it, and validating that the answer means anything. Each stage can silently invalidate everything after it, and only a handful of people in any laboratory can carry all three.

Setup is where studies quietly go wrong

Every code has its own input syntax, conventions and silent traps. A minor error in a configuration produces an incorrect result, a wasted allocation, or a failure discovered three weeks later. The cost is paid twice: in compute and in the time spent finding out.

Costly

Execution takes weeks before anything is usable

Preparation, equilibration, monitoring, restart, analysis. Weeks pass before the first result worth looking at, and the person doing it is usually the one you least want doing it.

Slow

Validation is skipped because it is expensive

Convergence, reference comparison, uncertainty. Everyone knows it should be done. Under deadline it gets shortened, and results get used anyway.

Risky

A few people carry all of it

Two or three people per laboratory can run these workflows end to end. Everything queues behind them, and when they leave, the protocols leave too.

Structural

02 / why now

AI opens a new era but not on its own.

For the first time, a scientist can describe an objective instead of writing input files. That is genuinely new. It is also not sufficient, and pretending otherwise is how a plausible answer becomes a wasted quarter of compute.

✓ What AI makes possible

Interaction in plain language

Describe a scientific objective without having mastered every tool it depends on.

Technical work, automated

Generate the code, prepare the files, chain the steps together without manual bookkeeping.

A scientist who is assisted

Explain, guide, analyse, and help interpret what the numbers actually mean.

✕ Where generic AI hits its limit

Broad knowledge, not expertise

Answers can be entirely plausible and still be scientifically wrong — the most expensive failure mode there is.

Results that aren't deterministic

The same request can lead to different methodological choices on different days. Science needs the opposite.

Workflows that are too complex

A real simulation demands a long chain of validations and expert decisions, not a single clever answer.

AI alone is not enough. It has to be framed by scientific workflows that are reliable, reproducible, and validated by experts.

03 / the approach

The reasoning of an AI, on methods built by experts.

MolAgent is not a chatbot that talks about chemistry. It runs the same simulation codes your group already runs, through methods a computational chemist designed, and it checks its own work along the way.

01 · AI reasoning

You describe, it plans

Understands the objective, decides on an approach, sequences the steps, explains what it is doing and why. You talk to it in ordinary language and it asks you when a decision is yours to make.

02 · Expert methods

Validated, structured, reproducible

The workflows are built and verified by computational chemists, grounded in the accepted practice of the field. The same request produces the same methodological choices, which is what makes a result defensible.

03 · Complete execution

Setup, execution and validation

All three stages, not just the middle one. It prepares, runs, monitors, evaluates its own results, and runs additional checks or calculations when it judges them necessary. You approve, adjust, or send it back. The agent installs alongside the simulation software you already run, so your structures and results stay inside your organisation.

Understands

Interprets the objective and identifies a suitable strategy, stating what would count as a usable result.

Prepares and runs

Configures the workflow, generates the inputs, and launches the calculations on your systems.

Monitors and checks

Watches execution, detects errors and instabilities, adapts the steps, and evaluates whether the result holds.

Analyses and proposes

Puts the results in context, concludes, states what remains uncertain, and proposes the next step.

For the expert Weeks of repetitive technical work delegated, and the studies you designed but never had time to run.

For the team without a specialist Simulation becomes usable, at a standard someone else can check.

HOW IT FITS

molagent · v0.1.0

YOU

Researcher

State the objective
Review the proposed approach
Approve, adjust, decide
Own the conclusion

↔

MOLAGENT

MolAgent CLI

Expert-built method library
Study design and setup
Input and job generation
Execution, monitoring, adaptation
Validation and reporting
Memory of your systems and conventions

↔

YOUR ENVIRONMENT

HPC · Workstation · Cloud

Your existing computing resources
Your installed simulation software
Your storage and your results

Setup & Job Generation

Validated inputs for LAMMPS, VASP, GROMACS, CP2K, ORCA and others, calibrated to your system and your group's conventions. Submission, monitoring, restart and scaling on your queues.

Study Design

Breaks an objective into a plan: what to compute, at what level of theory, in what order, at what cost. Presents the alternatives and the trade-offs.

Validation

Convergence testing, comparison against reference data, uncertainty, and a record of every decision taken. The agent evaluates its own results and runs further checks when it judges them necessary.

Potentials Specialised to Your Chemistry

Machine-learned interatomic potentials fine-tuned on your own system, with active learning and uncertainty quantification. Three orders of magnitude faster than the reference calculation, at an accuracy you can defend.

Screening & Exploration

Large sets of candidates with active selection of what to compute next. Structure prediction, polymorph ranking, miscibility, adsorption, barriers, transport.

Analysis & Interpretation

Structural, dynamic and electronic properties, mapped onto quantities an experimentalist measures. Figures, and an explanation of what they mean.

Laboratory Memory

Protocols, past results, what failed and why. Last year's study becomes an asset for this year's.

Inside Your Organisation

Installs where the work already happens. Nothing is uploaded.

04 / methods

Four scales, one search space.

Real studies move between levels of description: survey cheaply, look closer where the answer is uncertain, confirm at the highest accuracy the budget allows. Each level has its own codes, formats and silent traps, and a real project combines several.

01 · Electrons

Quantum

Describes electrons explicitly to predict structure, energy and reactivity. Reactions, electronic properties, small molecules.

VASP Quantum ESPRESSO CP2K ORCA PySCF

02 · Atoms

Molecular Dynamics

Follows the motion of atoms over time. Structure, diffusion, interactions, complex systems. Enhanced sampling and free-energy landscapes.

LAMMPS GROMACS NAMD OpenMM PLUMED

03 · Groups of atoms

Coarse-Grained

Groups atoms together to reach larger scales. Membranes, large assemblies, self-assembly, soft matter.

MARTINI DPD ESPResSo HOOMD-blue VOTCA

04 · AI + physics

Learned Potentials

Near-quantum accuracy at the speed of molecular dynamics, specialised to your chemistry with active learning and uncertainty quantification. Large-scale screening, long dynamics, complex systems.

MACE NequIP Allegro UMA GAP

05 · Composition & structure

Screening Pipelines

Automated exploration of composition and structure with active selection of the next calculation.

Materials Project OQMD AFLOW AiiDA Atomate

06 · Observables

Analysis & Validation

Properties mapped to measurable quantities, convergence and uncertainty reported with every number, reproducible notebooks and publication-ready figures.

MDAnalysis ASE pymatgen OVITO VMD

05 / roadmap

From working agents to a delivered product.

A working prototype is used by two scientists today, with five more ready for the beta.

Progress

Phase 3 of 6

Updated Q3 2026 · Next review Q1 2027

✓

2025

Complete

Foundation

Company incorporated. Architecture defined. First methods designed.

✓

H1 2026

Complete

Working agents

Agents running real studies end to end. Prototype with two scientists.

●

H2 2026

In progress

Alpha and pilots

Deployable version installed with partners. Two to three pilots. Five scientists ready for the beta.

H1 2027

Planned

First product

Deliverable product. First results published with a pilot partner.

H2 2027

Planned

Launch

General availability on annual licence. Seed round.

2028 →

Vision

Extension

Pharmaceutical, biomedical and energy. European and US expansion.

06 / markets

Where MolAgent creates value today.

Three customer archetypes with the same underlying pain and three different procurement paths, inside a market growing at roughly 12% a year and backed by major public investment in AI-for-science.

$4.4B

Computational chemistry
market in 2034

≈ 12%

Annual growth
over the period

Up from roughly $1.7B in 2026. Public momentum is building on top of it: over $5B committed to the US Genesis Mission across ~300 projects, €107M for Europe's RAISE pilot with materials science as a flagship area, and €2.5B for AI under France 2030.
Source: Fortune Business Insights

Universities & research groups

The need: accelerate projects, standardise methods, reduce dependence on two or three people. The value: the execution is automated and the group's practice is captured instead of walking out the door.

Computational chemistry Materials science Biomolecular labs

Computing centres & national laboratories

The need: open complex infrastructure to a wider set of users without multiplying human support. The value: more science per machine-hour, and a documented, auditable practice available to every user.

EuroHPC facilities Government labs AI factories

Industrial R&D

The need: accelerate discovery in-house, with confidential data and reliability requirements. The value: complete studies run inside the perimeter, with the validation attached so a decision can be taken on the result.

Pharma & biotech Energy Semiconductors

07 / team

Two complementary profiles.

Not an outsider's guess at what scientists need. The science comes from someone who has spent a career doing it by hand; the engineering from someone who has shipped AI systems at scale.