Portal de Periódicos da CAPES

Reaction: The Near Future of Artificial Intelligence in Materials Discovery

2018; Elsevier BV; Volume: 4; Issue: 6 Linguagem: Inglês

10.1016/j.chempr.2018.05.021

2451-9308

Rafael Gómez‐Bombarelli,

Topic Modeling

Rafael Gómez-Bombarelli is the Toyota Assistant Professor in Materials Processing in the Department of Materials Science and Engineering at MIT. He received a PhD in physical chemistry from the Universidad de Salamanca and did postdoctoral work at Heriot-Watt University, Harvard University, and Kyulux North America Inc. Since 2018, he has led an MIT computational research group that uses machine learning on first-principle simulations and experimental data to accelerate materials discovery. His team focuses on addressing the inverse design question from atomistic structure to device performance with an emphasis on molecular materials and soft matter. Rafael Gómez-Bombarelli is the Toyota Assistant Professor in Materials Processing in the Department of Materials Science and Engineering at MIT. He received a PhD in physical chemistry from the Universidad de Salamanca and did postdoctoral work at Heriot-Watt University, Harvard University, and Kyulux North America Inc. Since 2018, he has led an MIT computational research group that uses machine learning on first-principle simulations and experimental data to accelerate materials discovery. His team focuses on addressing the inverse design question from atomistic structure to device performance with an emphasis on molecular materials and soft matter. In their Catalysis piece, Ozin and Siler discuss the limits of algorithmic materials discovery. This maximalist question of machine supremacy, for lack of a better term, is incredibly interesting and will fill many pages in years to come—partly because it touches on what the human mind is on the metaphysical level. Nevertheless, in engineering any process, performance bottlenecks should be addressed first, and the question then becomes, is hypothesis generation the bottleneck of materials discovery? Are human minds in need of immediate computer assistance about new ideas to try? Lack of ingenuity is hardly the reason why new materials-based technologies typically have decade-long incubation periods. The back and forth between the inventive mind and physical reality is. Room-temperature superconductors and photosynthesis catalysts, to name a couple, have been on the materials holy grail list since at least 1995,1Bard A.J. Whitesides G.M. Zare R.N. McLafferty F.W. Holy grails of chemistry.Acc. Chem. Res. 1995; 28: 91Crossref Scopus (49) Google Scholar not because of a dearth of rationally designed, creative attempts at shifting paradigms but because of the complex feedback loop of ideation, experimentation, and knowledge integration. Can a computer program ever be written to surpass the creativity of a visionary, at their most inspired, on an open-ended task? Or the insight of an expert after years of meticulous study? A Mendeleev, a Lovelace, a Ramon y Cajal? Perhaps. Any time soon? Unlikely. Strong artificial intelligence (AI) (the conscious, sentient, synthetic mind) and artificial general intelligence (the one-size-fits-all AI) are holy grails and will remain so for some years. Exponentially growing hardware resources, algorithmic improvements in models and learning approaches, and a renewed wave of excitement and funding are rapidly pushing the field forward. The first stop is matching human performance at a given task. Then comes surpassing it. For some tasks, there is not even such a metric: synthetic speech cannot be more realistic than human utterances. On this voyage to the Ithaca of strong AI, I hope it is a long one—materials science and engineering have much to gain from the ever-broadening narrow AI of today. Materials science and technology are very human endeavors. There is room for the hard rules of physics and math, for the soft rules of accumulated empirical knowledge, gathered through many thousands of person years, and also for the spark of genius and imagining the so-far unimagined. Computers have already succeeded at the first and are making great progress on the second. The third can only come after, if at all. Computers are very good at rules. If the instructions of the game can be stated, or even if the rulebook can be inferred, a computer program is likely to outperform a human. It has happened in chess, jeopardy, and go, among others. Machine learning is, in a sense, all about learning the rules. The laws of physics can be written as a computer program, to some degree of approximation, and computational experiments can be run on a few interesting compounds. Or on a few million. That is precisely high-throughput screening: the robotization of physics-based simulations. A whole library of every potential material for an application can be screened in essentially no operating time. The most promising ones can then be singled out by hard rules of first-principle simulations. Nevertheless, high-throughput screening is not AI; it is mostly digital plumbing. After the inception of a new material comes synthesis, fabrication, and testing, and this is where the science becomes more craft. The rules of what procedures will work under which conditions are embedded in millions of scientific publications. More than any one scientist can dream to read—big data after all. In the last months, machine-learning approaches have learned the soft rules of inorganic and organic synthesis from the literature and showcased levels of performance that could not be imagined a few years ago.2Segler M.H.S. Preuss M. Waller M.P. Planning chemical syntheses with deep neural networks and symbolic AI.Nature. 2018; 555: 604-610Crossref PubMed Scopus (812) Google Scholar The synthesis gap is closing. But the negative data of what does not work are usually missing.3Raccuglia P. Elbert K.C. Adler P.D.F. Falk C. Wenny M.B. Mollo A. Zeller M. Friedler S.A. Schrier J. Norquist A.J. Machine-learning-assisted materials discovery using failed experiments.Nature. 2016; 533: 73-76Crossref PubMed Scopus (800) Google Scholar No matter the amount of existing literature for related systems, the synthesis and fabrication of a new material are fickle and generally involve some degree of tuning, of trial and error. The way scientific articles are written does not necessarily reflect the path the scientists took to reach their conclusions, and there is no bright red line between rational design and trial and error. How many attempts does it take to optimize a new synthesis? How many experimental conditions need to be tested before what seemed like a good idea is abandoned? Waves upon waves of graduate students—the unsung John Henrys of rational materials design—spend their formative years carrying out repetitive experiments and performing parameter sweeps or grid searches by hand. Good ideas can be abandoned too early, and bad ideas can persevere arbitrarily. A large fraction of laboratory work boils down to a combinatorial search over starting materials, operations, conditions, etc., to maximize a desired metric in the most replicable manner possible. That is essentially an algorithm, and algorithms are best left to computers. This connection of ideas to their physical implementation is the bottleneck of materials discovery, the current frontier of machine learning in materials discovery, and where the earlier gains will be seen: automated experiment design based on pre-existing results and the use of real-time feedback through automated laboratory instruments. Potentially liberating generations of scientists from menial, repetitive, slow tasks will not reduce their creativity because of overreliance. Writing did not make us forgetful, the printing press and cheap paper did not flood readers with confusing and harmful ideas, the radio did not ruin education, etc. The list goes on through the 20th and 21st centuries. Douglas Adams assuaged us about this.4Adams D. The salmon of doubt: Hitchhiking the galaxy one last time. Del Rey, 2002Google Scholar Involving AI in the mind-to-laboratory loop will empower the inventive and ingenious to be creative about materials as well as about intelligence. Their hypotheses will be quickly discarded or confirmed on the basis of the hard rules of physics and the capricious rules of laboratory work using millions of accumulated experiences and automated experimentation tools. Once that bottleneck is addressed, it might be time to pit machine against human ingenuity. Or perhaps there will be no holy grails left to worry about. As for the ethical and societal implications of AI—and there are many5Harari Y.N. Reboot for the AI revolution.Nature. 2017; 550: 324-327Crossref PubMed Scopus (35) Google Scholar—they are simultaneously more urgent and more mundane than mandatory genetic editing on unsuspecting subjects. Catalyst: New Materials Discovery: Machine-Enhanced Human CreativityOzin et al.ChemJune 14, 2018In BriefIs the limit of algorithmic materials discovery an illusion? Will machines built by humans match or surpass the ingenuity of humans to discover and synthesize an entirely new class of materials? Full-Text PDF Open Archive