Enterprise AI Analysis
Developing a complete AI-accelerated workflow for superconductor discovery
This paper presents a novel AI-accelerated workflow combining machine learning and DFT for rapid superconductor discovery, leading to the identification of 741 stable superconductors and experimental validation of two new compounds, Be2HfNb2 and Be2HfNb.
Accelerating Materials Innovation with AI
Our AI-driven workflow significantly reduces discovery time and cost, enabling breakthroughs in critical materials science. Key benefits include:
99.4%
True Negative Rate for Superconductivity
1.3M
Candidate Structures Screened
85.7k
Metals Screened from Databases
741
Stable Superconductors Identified
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
ML Model Performance
Workflow Efficiency
Experimental Validation
The core of our workflow relies on Bootstrapped Ensemble of Equivariant Graph Neural Networks (BEE-NET), trained to predict the Eliashberg spectral function. With the coarse phonon density of states (CPD) variant, we achieved a mean absolute error (MAE) as low as 0.87 K for superconducting transition temperature (Tc) predictions. Crucially, the model boasts a true-negative-rate of 99.4% for identifying non-superconductors, making it exceptionally efficient for large-scale screening where positive cases are rare.
We evaluated various loss functions, with the Earth Mover's Distance (EMD) function yielding the best overall performance in regression metrics for Tc. However, for classification tasks and high-throughput screening, the Mean Squared Error (MSE) loss function offered the highest true-negative-rate, minimizing redundant DFT calculations for non-superconducting candidates.
Our multi-stage, AI-accelerated discovery pipeline is designed for unparalleled efficiency. It reduced over 1.3 million candidate structures to just 741 dynamically and thermodynamically stable compounds with DFT-confirmed Tc > 5 K. The initial ML filters for band gap, formation energy, and predicted Tc enabled a rapid down-selection of candidates without any expensive DFT calculations, reducing 1.22 million candidates to 5.6k in mere seconds per material.
Subsequent stages involve coarse phonon density of states (PhDOS) screening and final high-accuracy DFT calculations for a²F(ω) only on the most promising candidates. This strategic integration of ML and DFT minimizes computational costs, allowing exploration of a vast materials space previously deemed impractical. The workflow effectively extrapolates beyond the bias of the training data, discovering materials with a significantly higher mean Tc than the training set.
A critical outcome of this study is the successful experimental synthesis and characterization of two previously unreported superconductors: Be2HfNb2 and Be2HfNb. These compounds were identified from our top predictions, selected based on high Tc, low formation energy, and known parent compounds (Be2Nb3, a conventional superconductor).
Both materials exhibit superconductivity, with onset Tc values of 3.18 K and 4.24 K respectively, confirmed through low-temperature transport and specific heat measurements. X-ray diffraction analysis further confirmed the successful introduction of Hf into the lattice, although some impurity phases were present, indicating the complexities of real-world synthesis. This experimental validation solidifies the practical utility of our AI-acceleraccelerated workflow for novel materials discovery.
0.87 K
MAE for Superconducting Critical Temperature Prediction
AI-Accelerated Discovery Workflow
Elemental Substitution & Database Query
→
ML Filtering (Eg, Ef, Tc, Eh)
→
DFT Optimization
→
Coarse Phonon DOS Screening
→
Final DFT Calculation (a²F(ω))
→
Experimental Verification
| Model Variant | Precision | TPR | TNR |
|---|---|---|---|
| CSO (MSE) | 0.84 | 0.47 | 0.98 |
| CPD (MSE) | 0.96 | 0.63 | 0.994 |
| CSO (EMD) | 0.80 | 0.63 | 0.97 |
| CPD (EMD) | 0.94 | 0.71 | 0.991 |
Experimental Validation: Be2Hf2Nb & Be2HfNb2
The workflow successfully predicted and led to the experimental synthesis and characterization of two novel superconductors: Be2Hf2Nb and Be2HfNb2. These materials were generated through elemental substitution and confirmed to exhibit superconductivity with onset Te values of 3.18 K and 4.24 K respectively. This validation closes the loop from theoretical prediction to real-world discovery, showcasing the power of integrated AI and DFT approaches.
Calculate Your AI-Driven Discovery ROI
Estimate the potential time and cost savings for your enterprise by integrating AI into your materials discovery pipeline.
Annual Savings
$0
Hours Reclaimed
0
Your AI Discovery Implementation Roadmap
Our proven three-phase approach ensures a seamless integration of AI into your existing research and development processes.
Phase 1: Data Strategy & Model Customization (4-6 Weeks)
We begin by assessing your current data infrastructure and materials objectives. Our team then customizes BEE-NET models to your specific material systems, leveraging your proprietary datasets alongside public resources.
Phase 2: Workflow Integration & Pilot (8-12 Weeks)
The AI-accelerated workflow is integrated into your R&D pipeline. We conduct a pilot project on a target material class, providing hands-on training for your scientists and refining the workflow based on initial results.
Phase 3: Scaling & Continuous Optimization (Ongoing)
Once the pilot is successful, we scale the workflow across your R&D operations. We provide continuous support and model updates, ensuring optimal performance and the ongoing discovery of novel materials.
Ready to Transform Your Research?
Connect with our AI strategists to explore how our specialized solutions can accelerate your materials discovery pipeline.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat