Stanford Physicists Create AI to Disrupt Laws of Nature
Innovative artificial intelligence program recreates the periodic table.
Posted Oct 11, 2018
Imagine being able to apply the power of artificial intelligence (AI) to invent novel materials that can potentially revolutionize many industries such as pharmaceuticals, biotech, electronics, plastics, semiconductors, glass, energy, nanotech, metal alloys, composite materials, ceramics, optics, and many more. In 2018, pioneering physicists at Stanford University in Palo Alto, California, announced in PNAS (Proceedings of the National Academy of Sciences of the United States of America) the creation of a new AI program (Atom2Vec) that was able to recreate the periodic table of elements — a milestone first step towards creating an AI that can discover new laws of nature, and invent novel materials and compounds . Atom2Vec was able to achieve this within just a “few hours,” versus the many centuries it took for humans . The way this was achieved was a cross-disciplinary AI approach — applying linguistic concepts to materials science.
Stanford physicists applied Zellig S. Harris’ hypothesis on the distributional structure of language to atoms instead of words. Harris’ linguistic concept puts forth the idea that basic classes of the entities of language can be grouped by distributional behavior because they tend to have similar distributional properties. To illustrate Harris’ idea, the word “aunt” is associated with “female” and “uncle” with “male.” A possible vector for “aunt” may be described as “aunt equals uncle minus male plus female. Drawing upon this linguistic analogy, the research team created Atom2Vec with concepts drawn from Google’s Word2Vec, a two-layer neural net for natural language parsing .
The physicists used “atom vectors as a basic input units for neural networks and other ML models designed and trained to predict materials properties.” Atom2Vec is based on converting basic data units into mathematical vectors which the AI program learns through recognizing patterns. For example, Atom2Vec was able to learn that sodium and potassium have similar properties based on the shared property of binding with chlorine.
This first iteration of Atom2Vec was based on unsupervised machine learning. This means that the algorithm was fed unlabeled input data without any corresponding output variables with the goal for the algorithm to learn inherent structure from the input data. For the next version, the team will leverage the breakthrough achieved in AI recreating the periodic table of elements to develop future treatments for cancer patients with a more supervised machine learning approach. The overall goal for Atom2Vec 2.0 is to identify the optimal antibodies with the least amount of toxicity and maximum efficacy to attack antigens on cancer cells. In efforts to find novel solutions for cancer immunotherapy treatments, researchers plan to map genes onto a mathematical vector to organize the more than 10 million antibodies in the human body. The future of Atom2Vec will span across disciplines from the realm of chemistry to biology, oncology, immunotherapy, and medicine.
Copyright © 2018 Cami Rosso All rights reserved.
1. Zhou, Quan; Tang, Peizhe; Liu, Shenxiu; Pan, Jinbo; Yan, Qimin; Zhang, Shou-Cheng. “Learning atoms for materials discovery.” Proceedings of the National Academy of Sciences of the United States of America (PNAS). June26, 2018.
2. Than, Ker. “Stanford AI recreates chemistry’s periodic table of elements.” Stanford News. June 25, 2018.
3. Cobley, Andrew. “Do you Word2Vec? Google’s neural-network bookworm.” The Register. 13 Oct 2017.