CAREER: Machine-Learning Assisted Process Systems Engineering: Hybrid modeling for process optimization, design and synthesis

职业：机器学习辅助过程系统工程：用于过程优化、设计和合成的混合建模

• 批准号: 1944678
• 负责人: Fani Boukouvala
• 金额: $ 54.68万
• 依托单位: Georgia Tech Research Corporation
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-02-01 至 2025-01-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1944678&HistoricalAwards=false
• 关键词: CAREER; Machine; Learning; Assisted; Process

Process Systems Engineering (PSE) has historically been advanced by using physics-based mathematical models and computer algorithms to design, optimize and control complex systems. Recent advances in the broad field of Data Science and Artificial Intelligence have led to a series of breakthroughs in the development of Machine Learning (ML) tools that can be used to derive mathematical models from large data sets. However, ML-based models can be limited in their ability to give insight into the physical origins of the system’s behavior and can result in poor predictions outside of the range of the original data set used to create the model. This CAREER project aims to develop hybrid modeling approaches that preserve what we already know about system behavior to make data-driven ML models more reliable leading to more accurate medical diagnoses, smarter autonomous vehicles, and safer chemical plants. Curriculum development activities are proposed aimed at introducing data science concepts into Chemical Engineering courses and high school statistics classes. Proposed outreach activities are aimed at increasing the number of female engineers in the field of PSE.The proposed methodology aims at developing algorithms that will enable the simultaneous training of modern ML models (i.e., Neural Networks and Gaussian Process Models) with physical constraints that are derived from discretization of first-principles based models. The proposed research will involve a systematic study of pre-processing and integration of data, identification of low-dimensional descriptive feature spaces, hybridization of ML models with first-principles based models and quantification of the uncertainty of hybrid model predictions. The specific research aims are: (1) Theoretically advancing mathematical techniques for training nonparametric ML models to satisfy first-principles based model predictions (hybrid modeling); (2) Quantification of the uncertainty of hybrid models in the presence of noisy and incomplete data sets; (3) Algorithmic development for mixed-integer nonlinear optimization problems for design and synthesis with embedded hybrid models. A series of case studies that include production of pharmaceuticals, polymers and chemicals will be used to develop a benchmarking library for testing various hybrid modeling architectures. The design and optimization of bioprocesses will be studied using hybrid modeling to connect gene-level control to macroscale process optimization. A set of hybrid modeling and optimization teaching modules, suitable for incorporation within existing core courses of the Chemical Engineering curriculum, will be developed and broadly disseminated. Outreach activities are proposed that are aimed at introducing data-driven decision-making to high-school students through statistics classes and increasing gender diversity in the field of PSE. A high-school teacher will be hosted by the Principal Investigator in collaboration with the Center for Education Integrating Science, Mathematics, and Computing (CEISMC)at Georgia Tech and the Georgia Intern Fellowships for Teachers (GIFT) program, which provides paid summer STEM internships in industry workplaces and University laboratories for K-12 science, mathematics, and technology teachers.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.

过程系统工程（PSE）在历史上一直是通过使用基于物理的数学模型和计算机算法来设计，优化和控制复杂系统。数据科学和人工智能广泛领域的最新进展导致了机器学习（ML）工具开发的一系列突破，这些工具可用于从大型数据集导出数学模型。然而，基于ML的模型在洞察系统行为的物理起源方面可能受到限制，并且可能导致用于创建模型的原始数据集范围之外的预测结果不佳。这个CAREER项目旨在开发混合建模方法，保留我们已经知道的系统行为，使数据驱动的ML模型更可靠，从而实现更准确的医疗诊断，更智能的自动驾驶汽车和更安全的化工厂。提出的课程开发活动旨在将数据科学概念引入化学工程课程和高中统计课程。拟议的推广活动旨在增加PSE领域的女性工程师数量。拟议的方法旨在开发能够同时训练现代ML模型的算法（即，神经网络和高斯过程模型），其具有从基于第一原理的模型的离散化导出的物理约束。拟议的研究将涉及数据预处理和集成的系统研究、低维描述性特征空间的识别、ML模型与基于第一性原理的模型的混合以及混合模型预测的不确定性的量化。具体的研究目标是：（1）从理论上推进训练非参数ML模型的数学技术，以满足基于第一性原理的模型预测（混合建模）;（2）在存在噪声和不完整数据集的情况下量化混合模型的不确定性;（3）嵌入式混合模型设计和综合的混合整数非线性优化问题的数学开发。一系列的案例研究，包括生产的药品，聚合物和化学品将被用来开发一个基准库测试各种混合建模架构。生物过程的设计和优化将使用混合建模来研究，将基因级控制与宏观过程优化相连接。一套混合建模和优化教学模块，适合纳入现有的化学工程课程的核心课程，将开发和广泛传播。拟议开展外联活动，旨在通过统计课向高中生介绍数据驱动的决策，并增加个人和企业领域的性别多样性。一名高中教师将由首席研究员与格鲁吉亚理工学院的教育整合科学，数学和计算中心（CEISMC）和格鲁吉亚教师实习奖学金（GIFT）计划合作主办，该计划提供带薪暑期STEM实习机会，在行业工作场所和大学实验室进行K-12科学，数学，该奖项反映了NSF的法定使命，并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，被认为值得支持。

项目成果

Physics-based Penalization for Hyperparameter Estimation in Gaussian Process Regression

• DOI: 10.1016/j.compchemeng.2023.108320
• 发表时间: 2023-06
• 期刊: Comput. Chem. Eng.
• 作者: Jinhyeun Kim;Christopher Luettgen;K. Paynabar;Fani Boukouvala

Perspectives on the integration between first-principles and data-driven modeling

第一性原理与数据驱动建模之间集成的观点

• DOI: 10.1016/j.compchemeng.2022.107898
• 发表时间: 2022
• 期刊: Computers & Chemical Engineering
• 影响因子: 4.3
• 作者: Bradley, William;Kim, Jinhyeun;Kilwein, Zachary;Blakely, Logan;Eydenberg, Michael;Jalvin, Jordan;Laird, Carl;Boukouvala, Fani
• 通讯作者: Boukouvala, Fani

Training Stiff Dynamic Process Models via Neural Differential Equations

通过神经微分方程训练刚性动态过程模型

• 发表时间: 2022
• 期刊: Computer aided chemical engineering
• 作者: Bradley, William;Gusmão, Gabriel;Medford, Andrew;Boukouvala, Fani
• 通讯作者: Boukouvala, Fani

Enabling global interpolation, derivative estimation and model identification from sparse multi-experiment time series data via neural ODEs

• DOI: 10.1016/j.engappai.2023.107611
• 发表时间: 2024-04
• 期刊: Eng. Appl. Artif. Intell.
• 作者: William Bradley;Ron Volkovinsky;Fani Boukouvala

Optimization with Neural Network Feasibility Surrogates: Formulations and Application to Security-Constrained Optimal Power Flow

使用神经网络可行性代理进行优化：安全约束最优潮流的公式和应用

• DOI: 10.3390/en16165913
• 发表时间: 2023
• 期刊: Energies
• 影响因子: 3.2
• 作者: Kilwein, Zachary;Jalving, Jordan;Eydenberg, Michael;Blakely, Logan;Skolfield, Kyle;Laird, Carl;Boukouvala, Fani
• 通讯作者: Boukouvala, Fani