Deep Learning Accelerated Inverse Design of Lab-Scale Energy Efficient Heterojunctions for Wide-Bandgap Devices

宽带隙器件实验室规模节能异质结的深度学习加速逆向设计

• 批准号: 2110033
• 负责人: Ming Hu
• 金额: $ 37.5万
• 依托单位: University of South Carolina at Columbia
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-08-01 至 2024-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2110033&HistoricalAwards=false
• 关键词: Deep; Learning; Accelerated; Inverse; Design

Efficient and reliable wide-bandgap high power transistors based on group III-nitrides (GaN, AlN, InN, etc.) are increasingly needed in today’s most demanding industries such as electric vehicles, data centers, radars, and consumer electronics. The functions of these systems rely heavily on the efficiency of removing excess heat from the devices’ active area, which is composed of function materials, substrates, and associated interfaces and heterojunctions, the dominant factor determining the overall interfacial thermal resistance (ITR) of the devices. Therefore, appropriate design and manufacturing of corresponding interfaces with minimized ITR are crucial to developing next-generation wide-bandgap devices. However, due to the immense search space of interfacial structures, it is impractical to evaluate all potential interfacial configurations using current trail-and-error experimental or computational approaches. One of the promising strategies is to use machine learning techniques, which are transforming the engineering field spanning from property predictions to inverse design. The overarching goal of this project is to develop novel deep neural network algorithms and workflow for the inverse design of lab-scale tailored interfacial structures to realize thermally efficient high-power wide-bandgap devices, along with experimental validation and demonstration. The success of this project will provide computational design tools and experimental fabrication protocols, not only to facilitate disruptive developments of key high-power electronic systems by breaking the bottleneck of thermal inefficiency issue, but also to speed up the material-to-industry processes. The project offers a unified platform to promote interdisciplinary collaborations spanning computational thermal science, experimental physics, and data science. The developed algorithms would benefit all engineers who study structure-device property relationships. This project will also increase public understanding and appreciation of machine learning for accelerating structure discovery and inspiring young researchers to pursue careers in STEM. Minority graduate students will get involved and trained in this interdisciplinary research project to strengthen high quality workforce in STEM.Aiming to address the main obstacles in the inverse design of heterojunctions for thermally efficient III-nitrides transistor devices, a set of key deep learning based techniques in the full inverse design pipeline will be developed: (1) deep neural network potentials will be developed, for calculating interatomic force constants to accurately and efficiently deal with large number of compositions with hundreds to thousands of nano-scale interfaces via nonequilibrium Green’s function method, which is not feasible for other traditional computational approaches. This will be facilitated by using frequency-resolved phonon transmission coefficient curves as the learning target in neural network training, which is the dominant factor in determining desired ITR across the functional interface or heterojunction and is unique for specific interface and provides more detailed hidden information of interfacial phonon transport. (2) powerful deep learning and spatial deep convolutional neural networks will be exploited in order to learn the features of phonon transmission curves and then unravel the complex, nonlinear, and usually implicit relationship between atomic structures of heterostructures and interfacial thermal resistance. (3) Genetic algorithms and the newest generative adversarial networks for the inverse design of hypothetical interfaces or heterojunctions will be developed. (4) A versatile and state-of-the-art technique, namely pulsed laser deposition, will be used to synthesize atomically thin films that have been theoretically proposed, with thickness approaching a monolayer, and their ITR will be validated. By combining computational thermal science with data science and experiment teams, this project will transform the study of complex interfacial thermal transport process using deep learning strategy and significantly accelerate the exploration process for optimal interfacial structures to achieve best thermal management performance and thus spur the practical implementation in semiconductor industry.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

基于III族氮化物（GaN、AlN、InN等）的高效可靠的宽带隙高功率晶体管如今，电动汽车、数据中心、雷达和消费电子产品等要求最高的行业越来越需要这种传感器。这些系统的功能在很大程度上依赖于从器件的有源区去除多余热量的效率，该有源区由功能材料、衬底以及相关的界面和异质结组成，这是决定器件的整体界面热阻（ITR）的主导因素。因此，适当的设计和制造相应的接口与最小的ITR是至关重要的开发下一代宽带隙器件。然而，由于界面结构的搜索空间巨大，使用当前的试错实验或计算方法来评估所有潜在的界面配置是不切实际的。其中一个很有前途的策略是使用机器学习技术，这将工程领域从属性预测转变为逆向设计。该项目的首要目标是开发新的深度神经网络算法和工作流程，用于实验室规模定制界面结构的逆向设计，以实现热效率高功率宽带隙器件，沿着实验验证和演示。该项目的成功将提供计算设计工具和实验制造协议，不仅通过打破热效率问题的瓶颈来促进关键大功率电子系统的颠覆性发展，而且还将加快材料到工业的过程。该项目提供了一个统一的平台，以促进跨学科的合作，包括计算热科学，实验物理和数据科学。所开发的算法将有利于所有研究结构-器件性能关系的工程师。该项目还将提高公众对机器学习的理解和欣赏，以加速结构发现，并激励年轻研究人员从事STEM职业。少数民族研究生将参与并培训这个跨学科的研究项目，以加强STEM领域的高素质劳动力。为了解决热效率III族氮化物晶体管器件异质结逆向设计的主要障碍，将开发一套完整逆向设计管道中基于深度学习的关键技术：（1）将开发深度神经网络势，用于通过非平衡绿色函数方法计算原子间力常数，以准确和有效地处理具有数百至数千个纳米尺度界面的大量组合物，这对于其它传统的计算方法是不可行的。这将通过使用频率分辨的声子传输系数曲线作为神经网络训练中的学习目标来促进，这是确定跨功能界面或异质结的期望ITR的主导因素，并且对于特定界面是唯一的，并且提供界面声子传输的更详细的隐藏信息。(2)将利用强大的深度学习和空间深度卷积神经网络来学习声子传输曲线的特征，然后解开异质结构的原子结构与界面热阻之间复杂的、非线性的且通常隐含的关系。(3)将开发用于假设界面或异质结的逆向设计的遗传算法和最新的生成对抗网络。(4)一个通用的和国家的最先进的技术，即脉冲激光沉积，将被用来合成原子薄膜，理论上已经提出，厚度接近单层，和他们的ITR将得到验证。通过将计算热科学与数据科学和实验团队相结合，该项目将利用深度学习策略改变复杂界面热输运过程的研究，并显著加快探索最佳界面结构的过程，以实现最佳热管理性能，从而推动半导体行业的实际实施。该奖项反映了NSF的法定使命，值得支持通过使用基金会的知识价值和更广泛的影响审查标准进行评估。

项目成果

Material transformers: deep learning language models for generative materials design

• DOI: 10.1088/2632-2153/acadcd
• 发表时间: 2023-03-01
• 期刊: MACHINE LEARNING-SCIENCE AND TECHNOLOGY
• 影响因子: 6.8
• 作者: Fu, Nihang;Wei, Lai;Hu, Jianjun
• 通讯作者: Hu, Jianjun

High-throughput computation of novel ternary B–C–N structures and carbon allotropes with electronic-level insights into superhard materials from machine learning

• DOI: 10.1039/d1ta07553e
• 发表时间: 2021
• 期刊: Journal of Materials Chemistry A
• 影响因子: 11.9
• 作者: Mohammed Al-Fahdi;T. Ouyang;Ming Hu

TCSP: a Template-Based Crystal Structure Prediction Algorithm for Materials Discovery

• DOI: 10.1021/acs.inorgchem.1c03879
• 发表时间: 2022-06-06
• 期刊: INORGANIC CHEMISTRY
• 影响因子: 4.6
• 作者: Wei, Lai;Fu, Nihang;Hu, Jianjun
• 通讯作者: Hu, Jianjun

Unlocking phonon properties of a large and diverse set of cubic crystals by indirect bottom-up machine learning approach

• DOI: 10.1038/s43246-023-00390-3
• 发表时间: 2023-08
• 期刊: Communications Materials
• 影响因子: 7.8
• 作者: Alejandro Rodriguez;Changpeng Lin;Chen Shen;Kunpeng Yuan;Mohammed Al-Fahdi;Xiaoliang Zhang;Hongbin Zhang;Ming Hu

Predicting lattice thermal conductivity from fundamental material properties using machine learning techniques

使用机器学习技术根据基本材料特性预测晶格热导率

• DOI: 10.1039/d2ta08721a
• 发表时间: 2023
• 期刊: Journal of Materials Chemistry A
• 影响因子: 11.9
• 作者: Qin, Guangzhao;Wei, Yi;Yu, Linfeng;Xu, Jinyuan;Ojih, Joshua;Rodriguez, Alejandro David;Wang, Huimin;Qin, Zhenzhen;Hu, Ming
• 通讯作者: Hu, Ming