Outlook: How I Learned to Love Machine Learning (A Personal Perspective on Machine Learning in Process Systems Engineering)
Outlook: How I Learned to Love Machine Learning (A Personal Perspective on Machine Learning in Process Systems Engineering)
复制标题
Outlook：我如何学会热爱机器学习（过程系统工程中机器学习的个人观点）
DOI：
10.1021/acs.iecr.3c01565
复制
发表时间：
2023
影响因子：
4.2
通讯作者：
Zavala, Victor M.
中科院分区：
文献类型：
--
作者：
Zavala, Victor M.
我一直在思考机器学习（ML）和相关领域（例如，人工智能，数字化和数据科学）如何改变，并将改变我们的过程系统工程（PSE）研究领域。在我写这篇评论的时候，ML领域继续爆炸式发展;到我写这篇评论的时候，将会有成千上万篇关于这个主题的新论文发表。这是我职业生涯中第一次见证这样的革命;见证这样的浪潮是有趣的，令人兴奋的，令人困惑的，有时甚至是可怕的。作为一名科学家，我继续想知道我的PSE技能如何适合并最好地用于这场ML革命。作为一名教育工作者，我一直想知道我应该为学生提供什么样的“旧”和“新”技能/知识的组合，以帮助他们充分利用机器学习革命。在这篇评论中，我提供了一个个人观点，关于我如何看待ML正在改变PSE，以及我认为我们（PSE社区）如何最好地使用ML并为ML做出贡献。在这样做的过程中，我将提供一个关于我如何“发现”（并学会爱上）ML的轶事视角，以及这在过去几年中如何改变了我的研究计划。在这种情况下，我将强调我的PSE培训促进我采用ML的方面。我还将分享一些关于ML教会我的令人惊讶/意想不到的事情的轶事，关于ML如何改变我感知世界的方式，以及ML如何为与不同研究领域的合作打开大门。我还将提供关于PSE社区正在进行的一些ML工作的观点;这将旨在提供一些见解，了解我认为ML和PSE如何最好地融合以产生切实的影响。作为其中的一部分，我将尝试将不同的ML范式置于背景中，并强调ML如何提供新的强大工具，帮助PSE研究人员做他们最擅长的事情：开发概念抽象。我还将试图论证为什么我认为在教育下一代PSE研究人员时，我们应该继续专注于教授第一原理建模和数学基础。我想强调的是，我不喜欢提供ML的精确定义，因为这是模糊的（从我的角度来看），并且快速发展并与其他领域融合。此外，我不想引发关于什么是ML与什么不是ML的不必要的讨论，因为这种判断本质上会受到观察者背景的偏见。例如，我认为主成分分析（PCA）是统计学的经典工具（而其他人认为它是ML工具），有些人喜欢说神经网络“只是”非线性回归模型。我认为，将技术放入垃圾箱是诱人的，但却是徒劳的，最终可能会阻碍创新。我将松散地使用ML这个术语，我通常指的是“现代数据科学”，而“现代”并不一定是指最近的，而是指“非经典”（从PSE的角度来看）。例如，深度神经网络、强化学习和Transformer模型是直到最近才在PSE中广泛采用的工具（尽管其中一些工具几十年前就被PSE研究人员探索过）。我还想强调的是，这篇评论文章并不是一篇科学文献综述，而是一个视角，我希望帮助读者快速了解ML领域，并帮助他们更批判性和选择性地浏览/判断文献。关于这类评论性论文的描述可以在参考文献1中找到。关于ML/AI和PSE（以及整个化学工程）之间交叉点的优秀评论和其他观点可以在参考文献2 - 6中找到。
I have been thinking a lot about how machine learning (ML) and related areas (eg, artificial intelligence, digitalization, and data science) are transforming and will transform our research field of process systems engineering (PSE). The ML field continues to explode as I write this commentary; there will be thousands of new papers published on the subject by the time I am done writing this. This is the first time in my career that I witness such a revolution; it has been fun, thrilling, confusing, and at times scary, to witness such a tidal wave. As a scientist, I continue to wonder about how my PSE skills fit and can be best used in this ML revolution. As an educator, I continue to wonder what mix of “old” and “new” skills/knowledge I should be providing my students to help them make the most of the ML revolution. In this commentary, I provide a personal perspective on how I see ML is transforming PSE and on how I think that we (the PSE community) can best use and contribute to ML. In doing so, I will provide an anecdotal perspective on how I “discovered”(and learned to love) ML and on how this has transformed my research program over the past few years. In this context, I will emphasize the aspects of my PSE training that facilitated my adoption of ML. I will also share some anecdotes on surprising/unexpected things that ML has taught me, on how ML has changed the way I perceive the world, and on how ML has opened doors to collaborations with diverse research fields. I will also provide a perspective on some of the ML work that the PSE community is conducting; this will aim to provide some insight into how I believe ML and PSE can best fuse to create tangible impact. As part of this, I will try to contextualize different ML paradigms and emphasize on how ML provides new and powerful tools that help PSE researchers do what they do best: develop conceptual abstractions. I will also try to argue why I think that we should continue to focus on teaching first-principles modeling and mathematical fundamentals when educating the next generation of PSE researchers. I would like to highlight that I prefer not to offer a precise definition of ML as this is blurry (from my perspective) and quickly evolving and converging with other fields. In addition, I do not want to trigger unnecessary discussions on what is ML vs what is not ML, because that judgment is inherently biased by the background of the observer. For example, I think that principal component analysis (PCA) is a classical tool of statistics (while others think that it is an ML tool) and some people like to say that neural networks are “just” nonlinear regression models. I think that aiming to fit techniques into bins is tempting but futile, and might ultimately hinder innovation. I will use the term ML loosely and by this I often mean “modern data science” and by “modern” I do not mean necessarily recent, but rather “non-classical”(from a PSE perspective). For instance, deep neural networks, reinforcement learning, and transformer models are tools that have not seen widespread adoption in PSE until recently (although some of these tools were explored decades ago by PSE researchers). I also would like to emphasize that this commentary paper is not intended to be a scientific literature review, but rather a perspective that I hope helps readers gain a quick understanding of the ML field and helps them navigate/judge the literature more critically and selectively. A description of what this type of commentary paper is supposed to be can be found in ref 1. Excellent reviews and other perspectives on the intersections between ML/AI and PSE (and chemical engineering at large) can be found in refs 2− 6.
登录
查看更多内容
影响因子：
4.3
作者：
Smith, Alexander;Zavala, Victor M.
通讯作者：
Zavala, Victor M.
SPT-NRTL：一种物理引导的机器学习模型，用于预测热力学一致的活动系数
DOI：
10.1016/j.fluid.2023.113731
发表时间：
2022
期刊：
ArXiv
影响因子：
--
作者：
Benedikt Winter;Clemens Winter;Timm Esper;J. Schilling;A. Bardow
通讯作者：
A. Bardow
针对成瘾进行优化：将产品责任概念扩展到设计有缺陷的社交媒体算法并克服通信规范法案
DOI：
10.2139/ssrn.3682048
发表时间：
2019
期刊：
SSRN Electronic Journal
影响因子：
--
作者：
Allison Zakon
通讯作者：
Allison Zakon
影响因子：
10.9
作者：
Akber Raza;Sharmistha Bardhan;Lihua Xu;Sharma S. R. K. C. Yamijala;C. Lian;Hyunah Kwon;Bryan M. Wong
通讯作者：
Akber Raza;Sharmistha Bardhan;Lihua Xu;Sharma S. R. K. C. Yamijala;C. Lian;Hyunah Kwon;Bryan M. Wong
影响因子：
1.9
作者：
Artur M. Schweidtmann;E. Esche;Asja Fischer;M. Kloft;J. Repke;S. Sager;A. Mitsos
通讯作者：
Artur M. Schweidtmann;E. Esche;Asja Fischer;M. Kloft;J. Repke;S. Sager;A. Mitsos