An Operating System for Materials Development

September 8, 2026

When we started PHIN, we were betting that uncertainty-aware machine learning models would be critical to actually replacing physical experiments. Over the past three years, we have proven that our models can indeed replace physical experiments. They can do this because instead of remaining fixed, our models autonomously self-improve. They generate new training data and fine-tune themselves to reach experimental accuracy across diverse disciplines including batteries, catalysts, water processing, and mining. We have moved simulation from a heuristic that narrows a design space to a method that labels which materials are best and which deserve a physical experiment.

This is a paradigm change in materials development, because most AI for materials applications can only reduce design spaces to around 1,000 candidates. Physically testing thousands of candidates for a single idea is impractical on the cost of the raw materials alone, let alone the labor and the equipment it takes. What we have been able to show is that we can take those 1,000 candidates and reduce them to 10 high-quality, validated materials solutions that actually deserve a physical experiment.

Managing the workstreams that whittle down millions of candidates and then running hundreds of digital experiments, however, is a different order of magnitude from what our first-generation product aimed to do. That product was meant for in-depth studies of tens of digital experiments. With the possibility of parallelizing hundreds of simulations and testing dozens of ideas at a time, we set out to reimagine how we could build an operating system for materials development.

We are excited to announce the initial release of our second-generation operating system, PhinOS, as we begin inviting external users onto the next-generation platform. For the next few weeks this is a closed testing period, and we are actively searching for the next crop of users to begin working on the platform.

The PhinOS command center, listing active threads, decisions awaiting review, and a fleet activity feed whose rows are tagged agent or human with the credits each has spent. — Every active thread, the decisions waiting on you, and live activity from people and agents alike.

What the platform entails

The foundation is still the same. You can use PHIN's models to run uncertainty-aware simulations that predict materials properties, rivaling the conclusions you can reach from physical experiments at 10% of the cost.
Project management features let you create research thrusts, projects, and threads. These facilitate scientific development by allowing you to recursively decompose a challenging research question into specific questions that a digital or physical experiment can answer.
Agents operate natively on the platform through two separate interfaces. The first is a traditional chat interface for analyzing data, creating workflows, and writing research documents. The second builds reusable, deterministic agentic workflows that perform structured knowledge work like reviewing papers and generating analysis code.
A native scientific document editor embeds your work directly, so every figure carries reproducible provenance for the data behind it and how it was generated.
High-throughput campaigns with Bayesian design-space optimization minimize the number of physical and digital experiments you run.
Multi-actor workflows compose agents, humans, and simulations into semi-autonomous research that puts human review at predetermined, high-leverage points.
A knowledge graph stores your multi-modal data, from papers to datasets to execution procedures, in one queryable, auditable place.
Governance surfaces predict costs, permission specific resources, share outcomes with different groups, and review the work of team members and agentic assistants.

A PhinOS research goal in the Iterate stage, showing a refuted root hypothesis, the supported hypothesis that refines it, and a proposed child awaiting a plan, each linked to the experiments behind it. — A research goal moves through Define, Research, Iterate, and Conclude. A refuted hypothesis spawns the alternative that replaces it.

Together, these form the substrate for increasing the productivity of scientists by an order of magnitude, by letting them build automations into their daily workflows and natively review and audit the findings. Concretely, here is what you can ask the platform to do:

Provision a simulation. "Provision and run an NVT MD simulation for Cu FCC at 800 K, 500 steps with the default PHIN calculator — then plot the RDF when it finishes."
Screen candidates. "Relax these perovskite candidates and rank them by formation energy, then show me the top five."
Run a bench experiment. "I am running solvothermal syntheses. I can set temperature between 120 and 220 °C in 5 °C steps, dwell 2–24 hours, and vary precursor concentration; I measure isolated yield and want to maximise it. I can run four per week — help me set up a campaign."
Query past experiments. "Which alloys have we measured above 900 K, what were the results, and who recorded them?"
Build a workflow. "Build a workflow that chains a relaxation into an elastic-constants run, with an approval gate before the second step."
Analyze data. "Find the 'Coin cell cycling — LiPF6 baseline' dataset already stored on the platform. Before anything runs, draft an analysis plan for me to review: plot discharge capacity against cycle number and compute the capacity fade rate per cycle. After I approve and launch the plan, determine whether the cell is fading and how fast, then prepare the resulting report for my review. Do not publish it internally until I approve the exact report version."

A PhinOS experiment graph for cation-vacancy formation energy, with 17 tasks across three rows showing completed, running, and queued states against a credit budget. — One experiment fanned out across three supercell sizes: 17 tasks, each priced and tracked from queued to complete.

The platform is the foundation of PHIN's next twelve months: unifying materials development and more broadly our companies mission of unifying engineering. It integrates the learnings from three years of deploying state-of-the-art simulation capabilities into experimental teams, and what it takes to make a step change in materials development productivity. It builds on our internal deployments of agentic tools like Claude, Codex, OpenClaw, Pi, and Hermes, and on the internal agentic projects we built for research and sales assistants. PhinOS combines it all into a substrate we plan to keep building on until we can run an entire materials R&D operation directly from PhinOS.

Rollout

So how can you begin using PhinOS?

First our proof points, two major projects we can discuss publicly. The first is a partnership with Pacific Northwest National Laboratory (PNNL) under an ARPA-E grant to develop new electrochemical catalysts 15x faster than traditionally capable. The work aims to develop bi-metallic catalysts for the conversion of CO2 to alcohol in less than a year, compared to the two or more decades it normally takes. The second is an internal initiative to develop a water desalination membrane that doubles the water produced at equivalent head pressures, halving the energy a desalination plant spends. Over the next six months we will show how we compress years of work into months, fully qualifying a material for end-to-end use.

While we are finishing internal testing with priliminary partners, we are preparing for the wider launch of PhinOS to enable doubling productivity of R&D scientists, managers, and directors and are actively demoing it with new interested parties. If you are running a materials program and want to accelerate your materials R&D, schedule a demo.