Elements Claw – Alibaba DAMO Academy's AI Agent for Superconducting Material Discovery

1. What is Elements Claw

Elements Claw is the industry's first AI agent for superconducting material discovery, jointly launched by Alibaba DAMO Academy, Renmin University of China, and the University of Chinese Academy of Sciences. It adopts a specialized-generalized fusion architecture, deeply coupling a 1-billion-parameter atomic foundation model, Elements, with a large language model (LLM) to achieve fully automated material screening—from literature retrieval and property prediction to experimental design. The agent completed screening of 2.4 million crystal structures in just 28 GPU hours, predicted 68,000 superconducting candidate materials, and successfully experimentally validated 4 new superconducting materials, marking a key breakthrough in AI's transition from "prediction" to "closed-loop validation" in functional material discovery.

Image source: Official article
Image source: official article

Technical Positioning and Domain: Elements Claw belongs to the AI for Science field, specifically focusing on superconducting material discovery in condensed matter physics and materials science. Unlike traditional high-throughput computing or single-property prediction models, it is positioned as an end-to-end material discovery agent system that integrates atomic-scale numerical computation with high-level semantic reasoning from LLMs, achieving for the first time a complete "AI design–experimental validation" closed loop in the superconductivity domain.

R&D Background: Led by Alibaba DAMO Academy's Intelligent Computing Lab, in collaboration with Renmin University's Gaoling School of Artificial Intelligence and the University of Chinese Academy of Sciences' School of Physical Sciences. The team has deep expertise in atomic-level pre-trained models (e.g., previous Uni-Mol, M3GNet) and LLM applications. The development was motivated by solving core pain points in traditional material discovery: long trial-and-error cycles (often decades), fragmented literature knowledge, and an exploding candidate space.

Core Value: Addresses three key challenges in superconducting material discovery: First, it accelerates screening speed by several orders of magnitude (28 hours vs. traditional decades); second, it improves prediction accuracy to experimentally verifiable levels (Tc prediction MAE of only 0.992K, superconductivity judgment AUC of 0.996); third, it enables continuous capability expansion through a self-evolution mechanism, autonomously creating new skills from literature without manual tool development. Its open database provides the global academic community with complete data resources for 2.4 million stable crystals.

Technical Features: Employs a specialized-generalized fusion architecture. The bottom layer uses the 1-billion-parameter atomic foundation model Elements, based on the EquiformerV2 equivariant architecture, for precise atomic-scale computation. The top layer uses an LLM for literature synthesis, task orchestration, and decision planning. It also possesses self-evolution capability: the agent can automatically extract new insights from literature and fine-tune the Elements model, enabling a closed-loop capability expansion of "discovering needs—creating tools—validating results."

2. Key Features

• Superconducting Critical Temperature Prediction: Based on the Elements-T model, it provides high-precision predictions of the superconducting critical temperature (Tc) for candidate materials, with a mean absolute error (MAE) controlled within 1K, significantly outperforming traditional DFT calculations and existing machine learning baselines. This feature directly supports property ranking in high-throughput screening, helping researchers quickly identify high-Tc candidates.

• Superconductivity Binary Classification: The Elements-C model determines whether a material is superconducting, achieving an AUC of 0.996, near-perfect classification. Serving as the first screening checkpoint, this feature rapidly eliminates non-superconducting materials from vast crystal structure databases, substantially reducing subsequent computational costs.

• Thermodynamic Stability Assessment: The Elements-E model evaluates the thermodynamic stability of candidate materials, including key indicators such as formation energy and convex hull distance, ensuring that predicted candidates are synthetically feasible under experimental conditions. This feature bridges prediction and experiment, avoiding the recommendation of thermodynamically unstable virtual structures.

• Crystal Structure Generation: The Elements-G model enables de novo crystal structure design, supporting motif-guided generation (e.g., specifying lattice type, atomic coordination environment) and exploration of novel phases. This feature can break through the limitations of known structure databases, generating unknown phases with potentially novel superconducting mechanisms.

• Intelligent Literature Synthesis: The LLM automatically retrieves, reads, and extracts superconducting knowledge from scientific literature, including reported Tc values, synthesis conditions, and crystal structure details, integrating scattered experimental knowledge into structured information to provide semantic support for screening and scheme design.

• Experimental Scheme Design: Combining the numerical computation results of the Elements model with the semantic reasoning capabilities of the LLM, it outputs executable synthesis pathways, heat treatment parameters, characterization plans, and validation strategies. For example, the Agent can recommend appropriate solid-state reaction temperatures, pressure conditions, and suggest specific steps for resistivity-temperature curve measurements and magnetic susceptibility tests.

• Self-Evolution and Skill Creation: The Agent can automatically extract new insights from literature and autonomously create new skill tools by fine-tuning the Elements model (e.g., Elements-C was created from literature through the self-evolution mechanism). This feature expands the agent's capability boundaries without human intervention, enabling continuous learning.

3. How to Use

1. Access the Open Database: Researchers can directly log in (link to be updated after official release) to access the complete data of 2.4 million stable crystals predicted by Claw, including crystal structure files (CIF format), predicted Tc values, stability indicators, etc. No registration is required for download, making it suitable for quick retrieval and secondary analysis.

2. Set Exploration Goals: In the system interface (or via API), input research requirements, such as "Screen for superconducting candidates with Tc greater than 4K in the Hf-Zr-Re ternary system." Natural language descriptions are supported, and the LLM will automatically parse the goal and break it down into screening parameters (element combinations, Tc thresholds, stability conditions, etc.).

3. Initiate Intelligent Screening: After submitting the task, Elements Claw automatically invokes Elements-T for high-throughput property prediction, while the LLM retrieves relevant literature and synthesizes existing evidence. The entire process requires no user intervention, and the system displays real-time progress in the task panel (number of structures, evaluation ratio, etc.).

4. Obtain Candidate List: Upon completion of screening, the system outputs a list of high-confidence candidate materials, along with predicted Tc values, stability indicators (formation energy, convex hull distance), synthesis feasibility assessments, and supporting literature evidence. Users can sort by various indicators and export the results as CSV or PDF.

5. Experimental Synthesis and Validation: Based on the synthesis plan recommended by the Agent (including precursor selection, reaction temperature/time, sintering conditions, etc.), prepare the target materials in the laboratory and conduct resistance-temperature curve measurements and magnetic susceptibility tests. The system also provides a comparative analysis function with predicted results to help verify model reliability.

Notes: The current system is primarily intended for academic research users, and API access requires permission application. For large-scale screening tasks, it is recommended to use GPU clusters to accelerate computation. The predicted results in the database are for reference only, and experimental validation remains a necessary step for final confirmation. The self-evolution feature is currently in beta and is only open to partner institutions.

4. Pros and Cons Analysis

5. Comparison of Similar Tools

Selection Recommendations: For research teams focused on superconducting material discovery, Elements Claw is currently the only tool offering an end-to-end agent workflow with experimental validation. Priority is recommended for using its open database and screening services. For scenarios requiring broad assessment of inorganic crystal stability, GNoME offers larger prediction scale (2.2 million stable crystals) but lacks experimental loop validation and agent decision-making capabilities. For research groups engaged in material generation (e.g., designing novel crystal structures), MatterGen's diffusion model approach offers greater flexibility, though subsequent property evaluation must be performed independently. MatBench is better suited as a benchmark platform for algorithm development rather than direct material discovery.

6. Editor's Take

The release of Elements Claw marks a significant leap in the AI for Science field, transitioning from "single prediction models" to "autonomous decision-making agents." Its technological innovation is reflected in three aspects: first, a specialized-generalized fusion architecture that deeply integrates a 1-billion-parameter atomic foundation model with an LLM, achieving seamless collaboration between numerical computation and semantic reasoning for the first time in materials discovery scenarios; second, a self-evolution mechanism, where the Agent can automatically extract knowledge from literature and create new skills—essentially a form of meta-learning that opens new pathways for continuous autonomous upgrades of AI systems; third, closed-loop validation, where all four initial AI-designed materials passed experimental verification, proving the feasibility of AI-driven materials discovery from theory to practice.

From a practical value standpoint, the efficiency advantage of screening 2.4 million structures in 28 hours directly addresses the core pain point of "combinatorial explosion" in materials science. The prediction accuracy, with an AUC of 0.996 and MAE of 0.992K, reaches a level usable for experiments, far surpassing traditional machine learning methods. The open database provides valuable shared resources for the global academic community, lowering research barriers. However, the current experimental validation sample size is only four, lacking statistical significance, and it primarily targets binary/ternary systems, with the generalization capability for quaternary and more complex systems requiring further testing. Additionally, the self-evolution mechanism is currently only available to partner institutions, and its actual effectiveness and stability need longer-term validation.

Target users include: experimental researchers in superconducting materials (for quickly obtaining candidate lists), computational materials scientists (as a high-throughput screening tool), and AI for Science algorithm developers (for studying specialized-generalized fusion architectures and self-evolution mechanisms). The future development potential is immense: as more experimental validation data accumulates, model accuracy will further improve; if the self-evolution mechanism is fully opened, it could foster an "AI scientist" ecosystem, enabling agents to autonomously iterate across multiple materials domains.

7. Application Scenarios

• Superconducting Material Discovery: Researchers input target element combinations and Tc thresholds, and Elements Claw automatically screens 2.4 million structures, outputting high-confidence candidate lists and synthesis schemes. It can quickly identify promising new superconductors in systems such as Hf-Zr-Re and La-Ce-B, reducing the traditional trial-and-error cycle from years to days.

• Solid-State Battery Electrolyte Development: Use Elements-E to evaluate the thermodynamic stability of ionic conductor materials, combined with Elements-T to predict ion mobility, screening candidate materials with high ionic conductivity and wide electrochemical windows. For example, searching for new garnet-type electrolytes in the Li-La-Zr-O system.

• Multiphase Catalyst Design: Generate crystal structures with specific active sites (e.g., single atoms, defect sites) via Elements-G, then have the LLM synthesize catalytic reaction mechanisms from the literature to recommend catalyst candidates with high selectivity and stability. Suitable for key reactions such as ammonia synthesis and CO2 reduction.

• Thermoelectric Material Exploration: Screen energy conversion materials that combine excellent Seebeck coefficients (high power factor) with low thermal conductivity. The Agent can automatically balance the components of the thermoelectric figure of merit (ZT) and recommend optimized formulations for materials such as Bi2Te3-based alloys and Half-Heusler compounds.

• Database Knowledge Completion: Leverage the LLM's literature synthesis capabilities to automatically extract known functional materials from scattered academic papers that are missing from mainstream databases (e.g., ICSD, COD), supplementing open databases and filling data gaps.

8. FAQ

Q: Does Elements Claw require local deployment? Can it be used directly online?
A: The current system provides online services. Researchers can directly visit https://science.damo-academy.com to query the prediction database without local deployment. For customized screening tasks, submissions can be made via the API interface, and GPU cluster acceleration is recommended for computation.

Q: How reliable are the predicted Tc values? What is the experimental validation success rate?
A: The MAE for Tc prediction is 0.992K, with an average error within 1K. All four initial AI-designed materials passed experimental validation, but the sample size is small, and the actual hit rate requires more testing. It is recommended to use predictions as a candidate list, with experimental validation remaining a necessary step for final confirmation.

Q: What element combinations are supported? Can quaternary and higher-order systems be handled?
A: The system is currently optimized for binary and ternary systems, covering most main-group and transition metal elements in the periodic table. The theoretical framework for quaternary and higher-order systems has been established, but screening accuracy may decrease. It is recommended to first validate with binary/ternary systems before expanding.

Q: How does the self-evolution mechanism work? Can regular users use it?
A: The Agent analyzes unmodeled knowledge from literature via LLM, automatically generates fine-tuning data, and updates the Elements model to create new skills (e.g., Elements-C). This feature is currently in beta and only open to partner institutions, with public access expected in future versions.

Q: What is the data format of the open database? How can it be downloaded?
A: The database provides crystal structure files in CIF format, along with CSV tables containing material ID, chemical formula, space group, predicted Tc, formation energy, convex hull distance, and other information. Batch downloads are supported, and queries can be made via REST API with specific conditions. Download link: https://science.damo-academy.com

Q: Compared to GNoME, what are the advantages of Elements Claw?
A: GNoME primarily predicts stability, while Elements Claw offers a full-process intelligent agent service from stability assessment, superconductivity determination, and Tc prediction to experimental design, with self-evolution capabilities and experimental closed-loop validation. In the vertical domain of superconducting material discovery, Elements Claw demonstrates significantly better accuracy and practicality.

Q: Is a background in machine learning required to use it?
A: No. The system provides a natural language interface. Researchers only need to input research goals (e.g., "Screen candidates with Tc>10K in the La-Ce-B system"), and the Agent will automatically complete complex computations and reasoning. However, a basic understanding of materials science is recommended for deeper interpretation of prediction results.

9. Project Address

• Open Database: https://science.damo-academy.com
• arXiv Technical Paper: https://arxiv.org/pdf/2604.23758