The Google Deep Mind AlphaZero program performance emphasizes the importance of a good understanding of fundamentals to enable efficient development of novel, green Analytical Chemistry approaches

• New developments in artificial intelligence tools are introduced • Fundamentals of analytical chemistry are required to properly apply emerging AI tools • Comprehensive assessment of analytical methods should include fundamental principles In December 2017, AlphaZero, an Artificial Intelligence (AI) program developed by Google’s DeepMind division, defeated Stockfish, the top-ranked chess engine at the time. Stockfish had established itself as the highest-performing chess program of its era, defeating not only the world’s best human players but also every other competing computer program it faced. Running on high-performance hardware, Stockfish was a finely tuned system capable of evaluating over 70 million positions per second to identify optimal chess moves. AlphaZero, by contrast, operated on only 80,000 positions per second. The fundamental difference between the two programs lay in their design philosophy. Stockfish was essentially designed to be a sophisticated database search engine, drawing on decades of accumulated chess knowledge and pre-established strategies. AlphaZero, on the other hand, was given only the fundamentals of chess and nothing more. From that bare foundation, AlphaZero was set to play against itself through a process of trial and error, progressively discovering and refining its own strategies; in other words, optimizing its own operation through simulation. After just nine hours of self-directed learning, it faced Stockfish in a 100-game match and went undefeated, winning 28 games and drawing the remaining 72. Careful analysis revealed why AlphaZero had such a decisive edge; the program had independently developed novel, previously unexplored chess strategies that fell entirely outside Stockfish’s database. In effect, AlphaZero surprised its opponent with moves that no existing record had anticipated. This outcome powerfully demonstrates a broader principle: AI systems built on sound fundamentals, and given the freedom to learn autonomously, can generate genuinely original insights and discover optimum solutions that have never before been considered. The implications for fields well beyond chess, including analytical chemistry, are considerable. For analytical chemistry scientists to take full advantage of AI’s capabilities, they must approach it much as AlphaZero was built – by grounding it in the fundamental principles of analytical chemistry rather than relying solely on literature and database knowledge. A scientist who understands the core principles governing the problem they are investigating, not just what has already been published, will be far better positioned to program an AI engine that can uncover genuinely new solutions. Admittedly, current AI engines are not yet flexible enough for convenient implementation by most scientists as they require good understanding of computer science fundamentals, but that is likely to change in the near future. What is worth nothing, however, is the idea that predicting structure and behavior of chemicals using fundamentals is not new to chemists. Since the introduction of the Schrödinger equation, computer-assisted calculation has been a cornerstone in the field. That equation describes the fundamental principles governing the movement and interaction of charged particles and forms the basis of Ab Initio programs, which calculate how electrons interact with atomic nuclei to form compounds and determine their structures. Using this approach, scientists have historically been limited to investigating relatively simple structures in vacuum conditions. The introduction of semiempirical methods, however, has expanded what is possible, enabling predictions involving more complex systems and processes, including those occurring in condensed phases. The eventual arrival of optical computers will push boundaries further still, as processing speeds increase dramatically. When this theoretical chemistry framework is combined with AI, the potential for accelerated progress grows considerably. From an analytical standpoint, integrating chemical properties with engineering multiphysics software deepens our understanding of system performance and opens the door to fundamentally new strategies. Furthermore, AI engines such as MuZero, recently introduced by DeepMind, or similar emerging tools, could assist scientists in identifying new fundamental principles as experiments are performed using online deep neural network learning. That said, the significance of any discoveries made through this method would still need to be critically evaluated by scientists with deep foundational knowledge. AI can illuminate new paths, but human expertise remains essential for interpreting what is found. The discussion above focuses on AI software built upon well-understood fundamentals, but the same strategic principles apply equally to research conducted by scientists through their own ingenuity, without any computer assistance. A scientist seeking truly original discoveries is more likely to find them by approaching a well-defined problem from a fundamental perspective rather than by surveying the published literature through databases searches. That said, familiarity with published work remains essential, not as the source of new ideas, but as the means by which scientists can recognize novelty when it appears. Advances reported in seemingly unrelated fields, when combined with strong foundational understanding of the problem at hand, can also spark ideas that lead to unexpected breakthroughs. This reliance on databases, whether by researchers or AI systems, brings with it an important caveat. The research available to scientists, including analytical scientists, does not consist solely of ideas that are correct and well-interpreted. A considerable portion of it contains “noise”: data that has been misinterpreted or that is simply wrong. Extracting reliable and useful information from that body of knowledge is often difficult without a deliberate strategy for filtering out the inaccuracies. This is true whether the reader is a human scientist or an AI engine trained on that same literature – a reminder that the quality of any insight, artificial or otherwise, depends heavily on the quality of the knowledge it draws from.