Our new AI model, SparksMatter, discovered CaMg₂Si₂ - a Ca-filled Mg-Si Zintl silicide - as a thermoelectric built only from stable, non-toxic, earth-abundant elements. Thermoelectrics are solid-state materials that convert heat directly into electricity (and electricity into cooling) with no moving parts, which makes them a key technology for harvesting the vast amounts of waste heat from engines, industry and electronics, and even fusion - but today's best ones rely on scarce or toxic elements like tellurium, lead and bismuth. This is why an earth-abundant, non-toxic candidate matters.
Our model's physical reasoning to come up with the design: Mg₂Si is a known earth-abundant thermoelectric but conducts heat too well; a heavy, weakly bound Ca cation in the Mg-Si framework should scatter phonons while keeping a moderate band gap. It generated 100 Ca-Mg-Si crystals with MatterGen, kept the six within 0.05 eV/atom of the convex hull (via MatterSim), and predicted band gaps of 0.44-0.57 eV and bulk moduli of 53-54 GPa (CGCNN). Follow-up lattice dynamics found three CaMg₂Si₂ polymorphs dynamically stable, with lattice thermal conductivity ≈6 W m⁻¹ K⁻¹ at 300 K and ≈2 at 1000 K. The AI proposed a chemical hypothesis first, then developed and applied a separate generative/physics pipeline to test it, and six surviving structures came back with that hypothesized composition. The video replays the reasoning process.
New paper out with Alireza Ghafarollahi in npj Computational Materials: SparksMatter, an AI that runs the full in-silico inorganic materials discovery cycle - ideation, planning, computational experimentation, critique and reporting - from a single plain-language query.
Why this matters: conventional ML models for materials are typically single-shot predictors or generators. They can predict a property or propose a structure, but they do not organize the next scientific step. Discovery, instead, works as a loop: hypothesize, test, critique, revise. The key advance here is a deep reasoning layer that incorporates physics to decide which scientific tool to use, how to interpret the result, and what to change as next step.
How it works: SparksMatter spawns a suite of AI agents - scientists, planners, coders, reviewers and critics - that write and execute code against materials tools: Materials Project retrieval; MatterGen for generative crystal design conditioned on chemistry, band gap or bulk modulus; MatterSim for relaxation and convex-hull stability; CGCNN for property prediction. Adversarial agents check each phase, the system revises its ideas, plans and code from execution results, documents its own limitations, and delivers a scientific report with a validation roadmap spanning DFT, phonons, transport, synthesis and characterization.
Two more discovery tasks SparksMatter ran autonomously:
1⃣Soft inorganic semiconductors: generated 112 structures conditioned on low stiffness and narrowed them to 59 candidates absent from the Materials Project after toxicity, stability, electronic, mechanical and database screening - bulk moduli 11-24 GPa, band gaps 0.4-3.9 eV.
2⃣Lead-free perovskites: filtered 154,879 Materials Project entries to 162 Pb-free ABO₃ candidates meeting structural, toxicity, stability and band-gap criteria, including LaAlO₃, BaZrO₃, SrSnO₃, CaTiO₃ and SrTiO₃.
Benchmark: the same three tasks were given to frontier reasoning models acting as expert materials scientists with web browsing but without the generation and prediction tools. A blinded LLM evaluator scored every response ten times on relevance, scientific soundness, novelty, and depth and rigor. SparksMatter scored highest in aggregate, with its strongest advantages in novelty and depth & rigor. Its main limitation was scientific soundness because much of the core screening still relied on surrogate models rather than direct first-principles or experimental validation - a gap the system identified, documented, and mapped out how to close.
Takeaway: putting generative models, executable code and physics-based simulators inside the reasoning loop lets an AI propose structures outside existing databases, test them, reject weak candidates, and say what evidence is still missing.
---
Code and data: https://github.com/lamm-mit/SparksMatter…
Paper (open access): https://nature.com/articles/s41524-026-02205-8…
[github.com GitHub - lamm-mit/SparksMatter](https://t.co/i7sEHQAMyH)
---
##### Comments
> **Ben Schulz @schulzb589** · [2026-08-18](https://x.com/schulzb589/status/2089706665549406506)
>
> Very cool. Maybe it can come up with a replacement for Ruthenium. Pretty rare catalyst with some unique properties.
> **Blue | Semis & AI Infra @BlueTradeIn** · [2026-08-18](https://x.com/BlueTradeIn/status/2089691585709793674)
>
> Very cool result. Abundance clears one gate; deployment still needs zT at realistic temperatures, low contact resistance and cycling stability. If the model can optimize all three together, this moves beyond a materials-screening demo.
[continuing from previous screenshot]
...assumptions accurately capture physical reality remains an empirical question. That is why we fabricated and tested the results.
We generated four actuator classes by crossing two stimuli - humidity and heat - with two responses: bending and twisting. The fourth, thermal twisting, required no new pipeline and no separate derivation within the framework. It emerged by composing a thermal stimulus module already validated in one case with a twisting module validated in another. The generated G-code produced the intended motion without manual redesign, and all four predictions fell within one experimental standard deviation of the measured response.
Why this matters:
1. For AI in science, this provides a physics-aware type system against which generative proposals can be checked - and rejected at the interface - before expensive simulation, fabrication, or experiment. It is roughly analogous to proof checking, but for the composition of physical mechanisms.
2. For engineering, the accessible design space can scale with a library of validated components rather than with the number of individually derived cases.
3. The mathematics, category theory, carries all the way into a physical object on a print bed. This points toward scientific knowledge as executable infrastructure: models that are not only described in papers, but typed, composable, verifiable, and able to compile into experiments.
Excellent work led by my student @leemmarom with @SkylarTibbits & @GioeleZardini.
Note from Claude Sonnet 5
Conclusion of Markus Buehler's X thread: describes an experiment generating four actuator classes (humidity/heat stimuli x bending/twisting responses) where the fourth class (thermal twisting) emerged automatically by composing two already-validated modules, with predictions matching experiment within one standard deviation. Argues this gives AI-for-science a category-theoretic 'physics-aware type system' analogous to proof checking, letting design space scale with a library of validated components. Credits student @leemmarom with @SkylarTibbits and @GioeleZardini.
ai for sciencematerials sciencecategory theorybioinspired engineeringtwittermit
Markus J. Buehler [verified] @ProfBuehlerMIT
Can we compile matter - for instance, a pine cone - and derive new active materials, end-to-end from observation to manufacturing? If physical systems can be formalized as composable mathematics, we can point AI that has been shown to resolve long-open mathematical problems at matter itself. Our new work turns bioinspired engineering from analogy into formal compilation: biology and mechanics become explicit, checkable, and executable, so AI reasoning can produce physical designs.
This is the first end-to-end demonstration in which a formally compositional multiscale model is carried from a biological hierarchy, through engineered design and fabrication specification, to executable manufacturing code - and then to a physically tested artifact.
Background:
Humans have long been inspired by biology to advance technology, but this has usually been an ad hoc process rather than a mathematically rigorous one. Natural materials such as pinecones achieve adaptive behavior through mechanisms organized across many scales. Engineering typically translates those mechanisms by analogy: identify a biological principle, build something inspired by it, and validate each new design as a separate case. This can produce remarkable results, but the knowledge does not readily compound. Instead, we represent each scale as a dynamical module with explicit states, stimuli, governing laws, and interfaces. Every [cut off]
Note from Claude Sonnet 5
X post by MIT professor Markus J. Buehler announcing new research on 'compiling matter' — formalizing biological/mechanical hierarchies (e.g. pinecones) as composable mathematics so AI can carry a design end-to-end from biological observation through fabrication specification to executable manufacturing code and a physically tested artifact, replacing ad-hoc bioinspired-engineering analogy with formal compilation.
ai for sciencematerials sciencebioinspired engineeringtwittermit
[continuing from previous screenshot]
...advance technology, but this has usually been an ad hoc process rather than a mathematically rigorous one. Natural materials such as pinecones achieve adaptive behavior through mechanisms organized across many scales. Engineering typically translates those mechanisms by analogy: identify a biological principle, build something inspired by it, and validate each new design as a separate case. This can produce remarkable results, but the knowledge does not readily compound. Instead, we represent each scale as a dynamical module with explicit states, stimuli, governing laws, and interfaces. Every scale-to-scale map must preserve the stimulus-response dynamics: evolve the fine-scale system and then map upward, or map upward first and then evolve. The two paths must agree. Because this condition is preserved under composition, locally valid interfaces remain consistent when assembled into the full hierarchy.
We then carry that structure into an engineered system, translate the target behavior into a verified fabrication specification, and compile it into G-code: the toolpaths, deposition sequence, temperatures, speeds, and other commands executed by a 3D printer. The intermediate translations are explicit, checkable, and executable rather than completed through an ad hoc handoff.
The formal guarantee is that given valid local models and interfaces, their composition remains valid. Whether those models and manufacturing assumptions accurately capture physical reality remains an empirical question. That is why we fabricated and tested the results.
Note from Claude Sonnet 5
Continuation of Markus Buehler's X post explaining the technical method: representing each biological scale as a dynamical module with explicit states/interfaces, requiring scale-to-scale maps to commute (evolve-then-map equals map-then-evolve), then compiling the composed model into verified fabrication G-code for a 3D printer, with physical fabrication and testing as the empirical check.
ai for sciencematerials sciencebioinspired engineeringtwittermit
Markus J. Buehl... ✓ @ProfBuehlerM... · 2h
What a time to be alive! We are entering the era of machines that discover and build. Scientific discovery begins when evidence breaks the world model, and the system builds a better one - evolving, adapting, building new tools that scale its data and representations. That was the core argument of my keynote "Superintelligence for Scientific Discovery: Multi-Agent Swarms and Large Reasoning Models" at the @BerkeleyRDI Agentic AI Summit 2026. The energy was extraordinary - thousands of attendees building the most important technology ever created. Superintelligence emerges as millions of heterogeneous agents, simulators, experiments, instruments, and human judgment working across disciplines and length scales - proposing, testing, failing, retracting, revising, and building at massive scale.
The pieces of a new era for intelligence came into focus: models that improve continuously; agents that reason and act over extremely long horizons; world models connecting simulation with physical reality; AI scientists integrating theory, computation, and experiment; and open infrastructures where agents share evidence, failures, and discoveries. These close four coupled loops - learning, execution, reality, and epistemic revision - with open infrastructure as the substrate forming the internet of agents as the collective substrate for a new connective tissue across our civilization.
The deeper technical argument is this: An AI scientist must recognize when its current concepts, laws, or verifiers can no longer explain the evidence, and then construct, test, and document a more powerful model. In my talk, I showed concrete examples of how we are building toward this across scales:
[cut off]
Note from Claude Sonnet 5
Long tweet by MIT professor Markus J. Buehler (likely Markus Buehler) about his keynote "Superintelligence for Scientific Discovery: Multi-Agent Swarms and Large Reasoning Models" at the Berkeley RDI Agentic AI Summit 2026, arguing superintelligence will emerge from swarms of agents doing science. Text continues past the visible screen and is cut off.
aisuperintelligencescientific discoveryagentic aitwitter
The pieces of a new era for intelligence came into focus: models that improve continuously; agents that reason and act over extremely long horizons; world models connecting simulation with physical reality; AI scientists integrating theory, computation, and experiment; and open infrastructures where agents share evidence, failures, and discoveries. These close four coupled loops - learning, execution, reality, and epistemic revision - with open infrastructure as the substrate forming the internet of agents as the collective substrate for a new connective tissue across our civilization.
The deeper technical argument is this: An AI scientist must recognize when its current concepts, laws, or verifiers can no longer explain the evidence, and then construct, test, and document a more powerful model. In my talk, I showed concrete examples of how we are building toward this across scales:
1 Graph-native large reasoning models make mechanisms, relationships, and abstractions compositional, compilable, and inspectable.
2 Adversarial Builder-Breaker agents generate new evidence, attack their own principles, and accept, reject, or retract model revisions.
3 Self-organizing swarms develop their own meta-reasoning structure through interaction. ScienceClaw × Infinite (arXiv:2603.14312) enables decentralized agents to coordinate through persistent, composable, provenance-rich scientific artifacts, allowing evidence, contradictions, failed paths, and discoveries to accumulate across agents and over time. We have obtained remarkable results such as new protein sequences with wet-lab
[cut off]
Note from Claude Sonnet 5
Continuation of the same tweet thread by Markus Buehler (MIT), listing numbered examples of AI-scientist infrastructure: graph-native reasoning models, adversarial builder-breaker agents, and self-organizing swarms coordinating via a system called ScienceClaw x Infinite, citing arXiv:2603.14312. Ends mid-sentence mentioning new protein sequences validated with wet-lab work, cut off before further detail.
aisuperintelligencescientific discoveryagentic aitwitter
1 Graph-native large reasoning models make mechanisms, relationships, and abstractions compositional, compilable, and inspectable.
2 Adversarial Builder-Breaker agents generate new evidence, attack their own principles, and accept, reject, or retract model revisions.
3 Self-organizing swarms develop their own meta-reasoning structure through interaction. ScienceClaw × Infinite (arXiv:2603.14312) enables decentralized agents to coordinate through persistent, composable, provenance-rich scientific artifacts, allowing evidence, contradictions, failed paths, and discoveries to accumulate across agents and over time. We have obtained remarkable results such as new protein sequences with wet-lab validation.
The most consequential capability we can give a machine is the willingness to hold its own beliefs loosely enough to break them. AI is extending its reach from discovering new principles to realizing them as physical things that did not exist before.
Thank you to @BerkeleyRDI @dawnsongtweets for organizing this event and to everyone whose questions, ideas, and conversations made this such an extraordinary gathering.
Note from Claude Sonnet 5
End of the same Markus Buehler tweet thread: closes the numbered list of AI-scientist capabilities, makes a general philosophical claim about machines revising their own beliefs, and thanks Berkeley RDI and Dawn Song for organizing the summit.
aisuperintelligencescientific discoveryagentic aitwitter
A transformer can learn not just the outcomes of dynamics, but the operator that executes the rules. To show this we trained a transformer on roughly 0.04% of a discrete rule space - 100 of 262,144 possible rules - and it learned to apply unseen rules from the same rule class. The model does not simply memorize specific rules. It learns the operator that maps a supplied rule plus an initial state, including unseen rules from this class, to the correct next state. This is relevant because it is a shift from “neural networks approximate dynamics” to “neural networks can learn to execute symbolic programs within a defined rule class”. The rule itself is supplied at inference time, as data, and the network has internalized how rules act, not which rules to apply. On previously unseen rules, the model achieves 98.5% perfect one-step forecasts and reconstructs governing rules with up to 96% functional accuracy.
Two results make this hold up under scrutiny. First, inductive bias decay. As we scaled training rule diversity, the correlation between functional inference accuracy and distance-from-nearest-training-rule collapsed to R² = 0.00. At the largest tested training-rule diversity, the model’s performance on a new rule shows no measurable dependence on how similar that rule is to anything it was trained on. The bias toward training data (the thing we worry most about in compositional generalization claims) is something we can measure decaying, and we find that at scale it is gone.
Second, an identifiability theory. We derive a closed-form expression for the number of rules consistent with a single observation. This reframes the inverse problem: failure to recover ground truth is not necessarily a model defect, but can be correct behavior when the data underdetermine the rule. The model is sampling the equivalence class; and identifiability is governed by coverage, not capacity.
The methodological move underneath both results is amortization. Classical work on rule inference (e.g. the Santa Fe EVCA program, evolutionary search over CA rule space) was per-instance: search the rule space for each new system. We replace that with a single forward pass of a transformer trained across many instantiations of the rule class. That is what makes symbolic rule inference scalable as a research direction rather than a curiosity.
We show that this works in a tightly constrained domain: binary, deterministic, local cellular automata on small grids. The locality-break experiment shows the model fails sharply when target systems violate its structural priors (which is itself a useful diagnostic, but it bounds the operator class). We don't yet know how this scales to multistate, higher-dimensional, or stochastic CA, or whether it transfers cleanly to non-CA systems whose coarse-grained dynamics admit local surrogates. The identifiability framework - what can be inferred from observation, given a hypothesis class - should transfer wherever finite local rules meet sparse data. The amortization argument transfers wherever per-instance symbolic search has been the bottleneck. Those are the pieces I expect to outlive the cellular automata setting.
Led by @JaimeBerkovich with Noah David, at @LAMM\_MIT. Out now in Advanced Science @AdvPortfolio (link to paper & code below).
---
Paper: https://advanced.onlinelibrary.wiley.com/doi/10.1002/advs.202511352…
GitHub: https://github.com/lamm-mit/AutomataGPT…
[advanced.onlinelibrary.wiley.com AutomataGPT: Transformer‐Based Forecasting and Ruleset Inference for Two‐Dimensional Cellular...](https://t.co/mhGnASunYW)
