Superconductors Team Recap Report

Overview

Over the past weeks, the Superconductors team has made significant strides in leveraging machine learning models for the discovery and classification of novel superconducting materials. Key activities include fine-tuning the MatterGen model for critical temperature (Tc) prediction, developing a Tc classification pipeline, evaluating generated materials, and drafting strategic documents for the GHOST initiative.

Key Contributions

1. MatterGen Tc Fine-Tuning

@will initiated the fine-tuning of MatterGen using the 3DSC dataset to conditionally generate crystal structures with specific Tc values. The fine-tuning process involved:

• Utilizing 4 A10 GPUs with adjusted training parameters.

• Generating 15 candidate structures conditioned on a Tc of 298.15 K.

• Achieving a high uniqueness bias, ensuring diverse chemical systems.

• Maintaining structural validity and low energy above the hull in generated samples.

2. MatterGen to Tc Classifier Development

User not found developed a simplified end-to-end pipeline integrating MatterGen with a Tc classifier to evaluate the superconducting properties of generated materials. Key highlights include:

• Generation of 32 materials per batch with subsequent Tc evaluation.

• Identification of potential high-Tc candidates, such as gen_8.cif with an estimated Tc of 113 K.

• Recognition of model calibration aligning with the rarity of superconductors.

3. Evaluation of Fine-Tuned Model

Building on the fine-tuned MatterGen model, User not found evaluated 400 candidate materials:

• Found 82 materials exhibiting some level of superconductivity.

• Discovered 52 new chemical systems, expanding the search landscape.

• Identified new superconducting families, though no room-temperature superconductors were found.

4. GHOST One Pager Drafting

@will drafted a one-pager for the GHOST initiative, outlining the problem statement, proposed solutions, technical differentiation, and market opportunities. The document emphasizes:

• Reducing computational expense in materials simulation.

• Utilizing ML models for efficient property-targeted sampling and evaluation.

• Democratizing access to deep tech resources and enabling high-throughput screening.

Common Themes

• Integration of ML Models: Consistent focus on leveraging machine learning, particularly MatterGen, for generating and evaluating superconducting materials.

• Model Fine-Tuning and Evaluation: Emphasis on fine-tuning models to improve accuracy and relevance in predictions, alongside rigorous evaluation of generated samples.

• Discovery of Novel Materials: Active pursuit of discovering new chemical systems and superconducting families, expanding the existing knowledge base.

• Strategic Planning: Development of strategic documents (e.g., GHOST one-pager) to align team efforts with broader goals and market opportunities.

Insights

• Model Performance: The fine-tuned MatterGen model demonstrates the ability to generate structurally valid and unique materials, albeit with challenges in accurately predicting high Tc values.

• Discovery Potential: Identification of new superconducting families highlights the model's capability to explore uncharted chemical spaces, offering avenues for further research.

• Strategic Direction: The GHOST initiative positions the team to capitalize on ML-driven materials discovery, emphasizing efficiency and democratization of resources.

Potential Open Problems

• High-Tc Prediction Accuracy: The fine-tuned model currently struggles to predict high Tc values accurately. Enhancing the model's understanding of factors influencing Tc is crucial.

• Stability of Generated Materials: Further work is needed to ensure the stability of generated materials under various conditions.

• Expansion Beyond 5 Elements: MatterGen's limitation to generating materials with up to 5 elements may restrict the discovery of complex superconductors requiring more elements.

Missing Connections

• Validation with Experimental Data: Integrating experimental validation processes to corroborate the predicted superconducting properties of generated materials.

• Enhanced Fine-Tuning Techniques: Exploring advanced fine-tuning methods or alternative datasets to improve the model's predictive capabilities for Tc.

• Collaboration with Domain Experts: Strengthening collaboration with materials scientists to interpret model outputs and guide experimental validations effectively.

Recommendations

1. Enhance Model Training:

   • Incorporate additional datasets, including those with higher Tc superconductors.

   • Experiment with different fine-tuning strategies to improve Tc prediction accuracy.

2. Strengthen Evaluation Pipelines:

   • Implement robust validation mechanisms, including collaboration with experimental labs for real-world testing.

   • Develop metrics to assess the stability and feasibility of generated materials comprehensively.

3. Expand Chemical Space Exploration:

   • Overcome the 5-element limitation by exploring or developing models capable of handling more complex compositions.

   • Focus on newly discovered superconducting families for targeted exploration and validation.

4. Strategic Collaboration:

   • Engage with materials science experts to interpret findings and prioritize promising candidates for experimental validation.

   • Foster interdisciplinary collaboration to bridge the gap between ML models and practical superconducting material discovery.

5. Leverage GHOST Initiative:

   • Utilize the strategic framework provided by the GHOST one-pager to streamline discovery processes and maximize impact.

   • Focus on democratizing access to ML-driven materials discovery tools, enabling broader participation from academia and industry.

Conclusion

The Superconductors team has made commendable progress in integrating machine learning models for the discovery and classification of superconducting materials. By fine-tuning MatterGen and developing a Tc classifier, the team has demonstrated the potential to uncover novel superconducting families. Addressing current challenges through enhanced model training, robust evaluation pipelines, and strategic collaborations will be pivotal in advancing towards the discovery of room-temperature superconductors.