Axiomatic_AI

We are excited to present at the NSF Institute for Artificial Intelligence and Fundamental Interactions (IAIFI) Summer Workshop this Thursday on our work at Axiomatic AI, building a formal library and benchmark for AI reasoning in quantum mechanics. We are formalizing 700+ quantum mechanics exercises in Lean, with the goal of building both a foundation for quantum mechanics in Lean and producing a large-scale benchmark for AI-assisted theorem proving and scientific reasoning in physics. Looking forward to sharing what we’ve learned and discussing how formal methods can help build more reliable AI for science! IAIFI Summer Workshop: https://lnkd.in/eq8mN8DT

https://lnkd.in/ghP8UXKf

At Axiomatic AI, we start from a simple premise: AI can generate a lot, but it can't yet be trusted. So we build the verification layer that checks whether what AI produces in engineering and science holds up, using both formal and informal methods to catch failure modes and feed corrections back to the agent for iterative refinement. If that problem interests you, come find us at #ICML to chat about the work or potential opportunities.

We are excited to announce the open-source release of AxProverBase, our simple automated prover for Lean 4. It is basically a baseline for evaluating theorem provers. Many agentic provers come with great numbers, but the setups are often different and tricky to reproduce, so you can't easily tell if something's better because of the new harness or because of an advancement of the underlying model. So we made the simplest workflow that works, for people to compare against. It's not supposed to be good, but to be the reference thing to beat (even though it already ranks reasonably well on PutnamBench compared to other open-source models). This work has been accepted at [ICML] Int'l Conference on Machine Learning 2026, where our team will present it in two weeks! Alongside this release, we are inviting everyone in the Lean community to participate in the beta of our GitHub integration for AxProverBase, which allows you to automatically run AxProverBase on sorrys within your own Lean projects to accelerate your development workflow. You can tag AxProver in a GitHub issue or a PR and it will fill in your proofs in the background while you are working on other things. AxProver will create commits, but you always control how it contributes to your formalization project. Using AxProver’s GitHub integration - Registration: Please register through GitHub here: https://lnkd.in/gBrhNykj - Invoke ax-prover: You can invoke AxProver on a single proof obligation or on a PR You can find more technical details, including our research paper and source code, at prover.axiomatic-ai.com. AxProverBase is fully open source, so you can also find instructions there if you want to try it out on your own hardware. We’re excited to hear your feedback on the integration and proving capabilities! Learn more and get started: Source code: https://lnkd.in/gsiGkJ-c Research paper: https://lnkd.in/g62-JK4A PutnamBench leaderboard: https://lnkd.in/geRbCFm5

Axiomatic x Marimo GPU: FDTD Directional couplers are among the most fundamental components in photonic integrated circuits — two waveguides placed close enough that light transfers between them by evanescent coupling. The "bent" variant (see https://lnkd.in/gDtHsA9W) curves both waveguides, widens the range of wavelengths where desired 50/50 coupling occurs. Designing them well requires understanding precisely how coupling depends on gap, radius, and length: a relationship that 3-D electromagnetic simulation resolves at the level precision manufacturing demands. That has historically meant commercial FDTD software, a capable workstation, and enough experience with electromagnetic simulation toolchains to set the problem up correctly. Our latest digital twin notebook changes each of those requirements. GPU-backed finite difference time domain (FDTD) simulation closes that gap. Built with Axiomatic Intelligence, we can deliver this now with molab’s accelerated hardware support running the open source JAX-based FDTDx solver (https://lnkd.in/g7hw8JNy) on GPUs. We take targeted spot-check simulations at the geometry points where coupled mode theory (CMT) and physical reality are most likely to diverge, and use them to calibrate the CMT parameters. The result is a digital twin model that pairs CMT's physical intuition and speed with FDTD's quantitative accuracy. The infrastructure barrier matters as much as the speed improvement. Answering a what-if question about a design variant — "what happens if I tighten the bend radius by 10%?" is now available in your browser. The physicist who wrote the paper and the engineer designing the circuit can ask the same computational questions from the same notebook. Explore the notebook at https://lnkd.in/ggG6nA-k, and read about molab with GPU here: https://lnkd.in/gUaxXQs7 #Photonics #DigitalTwin #JAX #GPU #HPC #MachineLearning #marimo #CoreWeave

Axiomatic’s Digital Twins x Marimo GPU Integration Digital twins provide a compelling promise for engineers: build a virtual replica of your physical system and close the loop with empirical data. This compresses the cycle between "what if" and "we know." For many working today, the simulations underpinning the virtual replica side of the story are a bottleneck – many of the tools we rely on were built in the era before widespread GPU adoption, and are often CPU-limited. At Axiomatic, we’ve been working with partners to address this bottleneck. Today, our preferred notebook system for creating these digital twins, Marimo, is unveiling GPU integration, unlocking big improvements in AI-generated and driven modeling. We use Marimo's reactive notebook model — git-native python files with a dataflow execution graph, deployable as interactive web applications — as it fulfills a vision of literate programming for the AI era. And now, GPU-backed computation turns that artifact from a document into a live instrument. As an example of the need for accelerated computation, running finite difference time domain (FDTD) simulation of electromagnetic propagation on a JAX-accelerated GPU moves a solver invocation from a multi-hour CPU job to seconds. Through this combination, the embarrassingly parallel work of engineering — thousands of design candidates swept across wavelengths and fabrication process corners — collapses to a single compiled call. Yield and robustness analysis, which requires sampling that full ensemble, becomes feasible within an interactive session rather than requiring a dedicated compute campaign. The result is a different class of tool. Marimo notebooks run as web applications. Backed by cloud GPU, anyone can open a URL and drive a 3-D electromagnetic simulator from a browser. AI generates the twin, the GPU makes it tractable, Marimo makes it interactive and distributable. What we used to call a research script is now a deployed engineering tool — and the gap between those two things has largely closed. We've been building digital twins for photonic and electromagnetic design on this stack. Explore our example notebooks at https://lnkd.in/ggG6nA-k, and read about molab with GPU here: https://lnkd.in/gUaxXQs7  #Photonics #DigitalTwin #JAX #GPU #HPC #MachineLearning #marimo #CoreWeave

For those interested in formal provers and the path to physics verification, you can meet our team at ICLR, where Leopoldo Sarra and team are presenting Axiomatic’s work at the VerifAI workshop on Sunday: https://lnkd.in/evTCXnag We’re also sharing AxProverBase, and our paper “A minimal agent for automated theorem proving” (https://lnkd.in/e_eDKmNp) which builds upon and simplifies our team’s effort from September. The main takeaway is simple: in formal reasoning, verification feedback plus well-designed framework can go a long way, delivering strong results without unnecessary complexity. Alongside the paper, we’re also open-sourcing the code and releasing a GitHub integration for Lean, so developers can experiment with the prover directly in practice: https://lnkd.in/epGGvD8J And of course, we’re hiring!

For those interested in formal provers and the path to physics verification, you can meet our team at ICLR, where Leopoldo Sarra and team are presenting Axiomatic’s work at the VerifAI workshop on Sunday: https://lnkd.in/evTCXnag We’re also sharing AxProverBase, and our paper “A minimal agent for automated theorem proving” (https://lnkd.in/e_eDKmNp) which builds upon and simplifies our team’s effort from September. The main takeaway is simple: in formal reasoning, verification feedback plus well-designed framework can go a long way, delivering strong results without unnecessary complexity. Alongside the paper, we’re also open-sourcing the code and releasing a GitHub integration for Lean, so developers can experiment with the prover directly in practice: https://lnkd.in/epGGvD8J And of course, we’re hiring!

𝐀𝐱-𝐌𝐞𝐚𝐬𝐮𝐫𝐞 𝐃𝐞𝐦𝐨 𝐢𝐬 𝐋𝐢𝐯𝐞 𝐚𝐭 𝐎𝐅𝐂 𝟐𝟎𝟐𝟔 Today at MPI Corporation 𝐁𝐨𝐨𝐭𝐡 𝟓𝟎𝟐, we’re showcasing 𝐀𝐱-𝐌𝐞𝐚𝐬𝐮𝐫𝐞 — a fully autonomous AI test system for photonic chip characterization. Photonic chip testing is still largely manual: engineers align probes, interpret instrument documentation, write custom scripts, and assemble analysis workflows to run standard measurements. Ax-Measure addresses this by connecting Axiomatic AI agents directly to lab instruments through a validated tool library and integrated physics model library. 𝐀𝐱-𝐌𝐞𝐚𝐬𝐮𝐫𝐞 𝐨𝐩𝐞𝐫𝐚𝐭𝐞𝐬 𝐚𝐮𝐭𝐨𝐧𝐨𝐦𝐨𝐮𝐬𝐥𝐲 𝐚𝐜𝐫𝐨𝐬𝐬 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥 𝐦𝐞𝐚𝐬𝐮𝐫𝐞𝐦𝐞𝐧𝐭 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 — from experiment planning to instrument execution, data acquisition, and physics-based parameter extraction — without manual scripting. 𝐊𝐞𝐲 𝐂𝐚𝐩𝐚𝐛𝐢𝐥𝐢𝐭𝐢𝐞𝐬: 𝐍𝐚𝐭𝐮𝐫𝐚𝐥 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐏𝐫𝐨𝐦𝐩𝐭𝐢𝐧𝐠 𝐟𝐨𝐫 𝐓𝐞𝐬𝐭 𝐂𝐨𝐧𝐭𝐫𝐨𝐥 — define measurement objectives and generate executable workflows directly from natural language prompts 𝐆𝐃𝐒-𝐭𝐨-𝐓𝐞𝐬𝐭 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐨𝐧 — extract device locations and probe coordinates directly from layout files, eliminating manual transcription 𝐀𝐠𝐞𝐧𝐭-𝐎𝐫𝐜𝐡𝐞𝐬𝐭𝐫𝐚𝐭𝐞𝐝 𝐌𝐞𝐚𝐬𝐮𝐫𝐞𝐦𝐞𝐧𝐭 — generate and execute complete experiment plans based on device type and test objectives 𝐏𝐡𝐲𝐬𝐢𝐜𝐬-𝐁𝐚𝐬𝐞𝐝 𝐀𝐧𝐚𝐥𝐲𝐬𝐢𝐬 — fit measured data to extract parameters such as Q-factor, group index, and propagation loss using built-in or user-provided models The system executes the full pipeline — from probing devices and sweeping spectra to parameter extraction — with each step traceable to raw data, tool use, and model assumptions. Ax-Measure enables measurement outputs to feed directly back into design and fabrication workflows. Check out video here https://lnkd.in/gx2NWUtR If you’re at OFC, come by 𝐌𝐏𝐈 𝐁𝐨𝐨𝐭𝐡 𝟓𝟎𝟐 in the South Hall to see a live demo on Lightium thin-film Lithium Niobate wafer We’re also running a scheduled demo at 2 PM tomorrow. #OFC2026 #Photonics #TestAutomation #AIforAutomation

This week our co-founder Dirk Englund and Axiomatic team member Kevin Schädler — together with colleagues at MIT, MITRE, Sandia, and the University of Arizona — published "𝗡𝗮𝗻𝗼𝗽𝗵𝗼𝘁𝗼𝗻𝗶𝗰 𝘄𝗮𝘃𝗲𝗴𝘂𝗶𝗱𝗲 𝗰𝗵𝗶𝗽-𝘁𝗼-𝘄𝗼𝗿𝗹𝗱 𝗯𝗲𝗮𝗺 𝘀𝗰𝗮𝗻𝗻𝗶𝗻𝗴" in 𝘕𝘢𝘵𝘶𝘳𝘦 One of the hardest problems in photonics is deceptively simple to state: how do you get light off a chip and into the world? The device is called a 𝗽𝗵𝗼𝘁𝗼𝗻𝗶𝗰 𝘀𝗸𝗶-𝗷𝘂𝗺𝗽. It's a nanoscale waveguide monolithically integrated on a piezoelectric cantilever, fabricated in a 200 mm CMOS foundry process. Differential stress in the thin-film layer stack causes the cantilever to passively curl ≈90° out of plane — turning what would normally be a fabrication artifact into a feature. The result is a chip-surface scanner that emits a sub-µm, broadband, diffraction-limited beam directly into free space, with kHz-rate mechanical resonances and Q > 10 000. Key results: • >50× higher footprint-adjusted spot rate than state-of-the-art MEMS mirrors • Full-color image and video projection demonstrated • Single-photon initialization and readout of silicon-vacancy (SiV) centers in diamond Fault-tolerant quantum computers will require optical control of thousands to millions of qubits, across distinct wavelength transitions spanning UV to near-infrared. The ski-jump platform operates across exactly that range, on a single wafer process. This is the kind of work that matters to us at Axiomatic_AI 𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗔𝗜 𝘁𝗼𝗼𝗹𝘀 𝗳𝗼𝗿 𝘀𝗰𝗶𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 depend on tight coupling between the digital and physical worlds — and they depend on hardware precise enough to close that loop. Dirk is also a co-author this year on work using AI agents to automate photonic integrated circuit design: the same people building the hardware are actively building the AI tools to design its successors. Congratulations to Dirk, Kevin, and the entire team! If you're at OFC Conference, in LA this week, we'd welcome the conversation. 🔗 https://lnkd.in/gMxqj3YK #Photonics #QuantumTechnology #PhotonicIntegratedCircuits #NatureJournal #AxiomaticAI