Deep Principle (@DeepPrinciple) header
Deep Principle (@DeepPrinciple) profile picture

Unlocking new chemical reactions and materials with AI.

Joined December 2024
We’re introducing Mira Science, a general-purpose computing platform for scientific research. Mira Science turns research methods into programs that can be composed, run, inspected, and modified. Researchers can orchestrate teams of AI scientists, connect professional tools and cloud compute, and trace every result back to its data, code, execution history, and review. Paired with Hydrogen-1.0, @deepprinciple’s open-weight model for scientific research, Mira Science achieves a 48.3% pass rate on Frontier Science Research, compared with 46.7% for Codex + GPT-6 Astra and 45.0% for Claude Code + Opus 5.5. We believe AI scientists should be inspectable, extensible, and able to work in the same environment as researchers, their tools, data, and compute—not remain isolated inside a chat interface. Mira Science is now available on desktop. Register and download: mirascience.app/?utm_source=…
49
206
17
3,881
2,190,569
Start exploring Hydrogen-1.0. Download the open weights to adapt and evaluate the model for your own research tasks, or try it directly in MIRA.For training methods, detailed evaluations, and application examples, explore our technical blog. → Model weights: huggingface.co/deep-principl… → Try Hydrogen-1.0: mira.deepprinciple.com/ → Technical blog: blog.deepprinciple.com/intro… We look forward to seeing what you explore.
6
231
In our metal-complex search evaluation, Hydrogen-1.0 used feedback from previous recommendations to search for candidates with larger HOMO–LUMO gaps. A concrete example of using successive evaluation results to guide materials exploration toward a defined target.
5
237
Hydrogen-1.0 helps researchers develop procedures and specify individual operations. Lab Compile then checks equipment capabilities, prerequisites, and execution order in a virtual environment. One concrete example: flagging an instruction to aspirate liquid before a fresh pipette tip has been fitted. By checking these details virtually, researchers can identify and revise execution issues before moving to the lab.
6
262
Hydrogen-1.0 helps researchers investigate material properties, compare reaction pathways, and identify promising directions for experiments through calculations and simulations. In our polyethylene example, it builds a molecular model and writes simulation code. When a density calculation produces an abnormal result, it helps trace the issue to double-counting, correct the code, and rerun the calculation. These calculations give researchers a clearer basis for deciding what to test next.
4
313
In our evaluations, Hydrogen-1.0 achieved: - 38.3% problem-solving success on FrontierScience Research (FSR), 5.0 percentage points above GPT-6 Astra. - 93.0% accuracy on GPQA Diamond, close to Kimi K3’s 93.5%. These benchmarks examine two essential capabilities for scientific work: reasoning through unfamiliar, multi-step problems and answering challenging questions in physics, chemistry, and biology.
2
15
1,347
Introducing Hydrogen-1.0, Deep Principle’s first general-purpose large AI model, built to support scientific discovery. Hydrogen-1.0 helps researchers reason through complex scientific questions, carry out calculations and simulations, turn research ideas into experimental plans, and use feedback to guide further exploration. Today, we’re releasing its model weights for researchers and developers to adapt and evaluate on their own scientific tasks. Hydrogen-1.0 is also available in our AI Scientist product, MIRA. Explore the benchmarks, research examples, and ways to get started in the posts below. Building intelligence for scientific discovery.
31
54
12
710
1,311,776
We just dropped Materials Property Axiom (MPA). The problem: most materials foundation models crush it on computed properties, but choke on real experimental data (boiling points, solubility, flammability, etc.). MPA fixes this with physics-aligned transfer learning: • 64M structures → 3D pre-training • Thermochemistry + electronics + solvation → mid-training alignment • Hybrid attention/additive readout → experimental fine-tuning Results: ~10% better across 40 benchmarks, up to 51% on hard OOD tasks. The rule: match the physics, not just the data. Live at sciclaw.ai 🔗 blog.deepprinciple.com/intro…
13
33
392
52,592
The article highlights the "Scientific Discovery Evaluation" (SDE) benchmark, led by Deep Principle. Unlike traditional evaluations, SDE does not consist of isolated, difficult questions. Instead, it distills 43 research scenarios and a total of 1,125 interrelated tasks from 8 ongoing, real-world research projects with unpublished data.
17
134
1,664
244,050
The article opens by pointing out that the scientific community has long questioned traditional question-and-answer evaluation benchmarks (such as HLE). These tests often focus on obscure or even trivial knowledge points, making it difficult to measure an AI's ability to conduct real scientific research.
4
393
Recently, Science published a feature article titled "How will we know if AI is smart enough to do science?". The article focuses on the latest breakthroughs in evaluation systems for artificial intelligence's scientific capabilities, providing an in-depth analysis of the cutting-edge explorations by institutions such as Deep Principle, OpenAI, and FutureHouse in developing benchmarks for scientific intelligence.
4
2
10
1,827
Thrilled to be on the cover of JACS! @ACSPublications LLM‑EO (LLMs + evolutionary optimization) designs transition‑metal complexes, surfacing top candidates and novel ligands from a 1.37M space. pubs.acs.org/doi/10.1021/jac… From cover to frontier—Deep Principle keeps driving GenAI deeper into materials. More breakthroughs ahead. #AI4Science #MyACSCover
2
10
945