Exploring diamondlike lattice thermal conductivity crystals via feature-based transfer learning
At a Glance
Section titled âAt a Glanceâ| Metadata | Details |
|---|---|
| Publication Date | 2021-05-10 |
| Journal | Physical Review Materials |
| Authors | Shenghong Ju, Ryo Yoshida, Chang Liu, Stephen Wu, Kenta Hongo |
| Institutions | Japan Advanced Institute of Science and Technology, The University of Tokyo |
| Citations | 50 |
| Analysis | Full AI Review Included |
Executive Summary
Section titled âExecutive Summaryâ- Methodological Breakthrough: Developed a feature-based Transfer Learning (TL) model to bridge the gap between readily available âbig dataâ (harmonic phonon properties, P3) and scarce âsmall dataâ (lattice thermal conductivity, ÎșL).
- Extrapolative Prediction: The TL model successfully achieved âextrapolative prediction,â accurately identifying ultrahigh ÎșL materials (1000-3000 W/m-1K-1) despite being trained only on materials with ÎșL less than 370 W/m-1K-1.
- Novel Candidates Identified: Screened over 60,000 compounds, confirming 14 novel crystals with ÎșL comparable to or exceeding diamond, including cubic BAs (3411 W/m-1K-1) and wurtzite BAs (2947 W/m-1K-1).
- Physical Insight on Hardness: Demonstrated that while superhard materials often possess high elastic constants and phonon group velocity, superhardness alone does not guarantee high ÎșL; the phonon relaxation time (related to P3) is the critical limiting factor.
- Key Descriptors for Anharmonicity: Identified average/maximum dipole polarizability and van der Waals (VdW) radius as the leading compositional descriptors correlating qualitatively with anharmonic phonon scattering.
- New Allotropes: The screening identified promising allotropes of known high-ÎșL materials (e.g., lonsdaleite, hexagonal diamonds, and various BC2N phases) that warrant further synthesis investigation.
Technical Specifications
Section titled âTechnical Specificationsâ| Parameter | Value | Unit | Context |
|---|---|---|---|
| Maximum Calculated ÎșL | 3411 | W/m-1K-1 | Cubic Boron Arsenide (BAs) |
| Diamond ÎșL (Calculated) | 3048 | W/m-1K-1 | Iterative Boltzmann Transport Equation (IBTE) solution |
| Cubic BN ÎșL (Calculated) | 1876 | W/m-1K-1 | IBTE solution |
| Total Compounds Screened | >60,000 | N/A | Materials Project database |
| P3 Training Data Set Size | 320 | Crystals | Harmonic three-phonon scattering phase space (Feature Property) |
| ÎșL Training Data Set Size | 45 | Crystals | Lattice thermal conductivity (Target Property) |
| Training Data Maximum ÎșL | 370 | W/m-1K-1 | Upper limit of the small training data set |
| Minimum P3 Value | 0.6397 x 10-4 | cm | Cubic BAs (lowest scattering phase space) |
| NN Input Descriptors | 290 | N/A | Compositional features based on 58 elemental properties |
| NN Hidden Layers | 4 to 6 | N/A | Randomly generated structure for pre-training |
| DFT Energy Convergence | 10-8 | eV | Optimization threshold for lattice parameters |
Key Methodologies
Section titled âKey Methodologiesâ- Data Preparation: Collected crystal structures from Materials Project and harmonic force data for 320 materials from phonon databases. Calculated the three-phonon scattering phase space (P3) as the âbig dataâ feature property.
- Feature Engineering: Generated 290 compositional descriptors for each material based on 58 elemental-level properties (using XenonPy).
- Pre-Training (P3 Model): A fully-connected pyramid Neural Network (NN) was pre-trained using the 290 compositional descriptors to predict the P3 values of the 320 materials.
- Transfer Learning Implementation: The subnetwork of the best pre-trained NN model (excluding the output layer) was repurposed as a feature extractor, transferring the learned structure-property relationships.
- Target Training (ÎșL Model): A Random Forest model (200 trees) was trained using the 10-dimensional descriptors extracted from the transferred NN subnetwork and the âsmall dataâ set of 45 known ÎșL values.
- High-Throughput Screening: The trained Transfer Learning model was applied to screen the entire >60,000 compound database, prioritizing candidates with the lowest predicted P3 values.
- First-Principles Validation: The top-14 candidates were validated using rigorous anharmonic lattice dynamics calculations, solving the Boltzmann Transport Equation iteratively (IBTE) to ensure accurate ÎșL prediction, especially for ultrahigh conductivity materials.
Commercial Applications
Section titled âCommercial Applicationsâ- High-Density Thermal Management: Provides validated alternatives to diamond for use as heat spreaders and sinks in modern microelectronic devices and high-power density components (e.g., CPUs, GPUs, and power modules).
- Laser Diode and Power Electronics: Materials like cubic BAs and various BN/BC2N allotropes offer superior ÎșL for managing thermal loads in high-performance laser systems and power switching devices.
- High-Temperature Stability: Identification of carbon nitrides (C3N4, BC2N) and boron compounds that are thermally and chemically more stable than diamond, mitigating thermal damage via oxidation or graphitization at high operating temperatures.
- Composite Materials and TIMs: The identified high-ÎșL insulators are ideal candidates for use as electrically insulating fillers in advanced Thermal Interface Materials (TIMs) and composites, crucial for preventing electrical leakage in complex integrated structures.
- Materials Discovery Acceleration: The feature-based transfer learning methodology provides a rapid, scalable screening tool for discovering materials with complex, higher-order properties (like ÎșL), significantly reducing the reliance on time-consuming first-principles calculations.
View Original Abstract
Ultrahigh lattice thermal conductivity materials hold great importance since\nthey play a critical role in the thermal management of electronic and optical\ndevices. Models using machine learning can search for materials with\noutstanding higher-order properties like thermal conductivity. However, the\nlack of sufficient data to train a model is a serious hurdle. Herein we show\nthat big data can complement small data for accurate predictions when\nlower-order feature properties available in big data are selected properly and\napplied to transfer learning. The connection between the crystal information\nand thermal conductivity is directly built with a neural network by\ntransferring descriptors acquired through a pre-trained model for the feature\nproperty. Successful transfer learning shows the ability of extrapolative\nprediction and reveals descriptors for lattice anharmonicity. Transfer learning\nis employed to screen over 60000 compounds to identify novel crystals that can\nserve as alternatives to diamond.\n