日本語 
What if the very material we need is already there, but we just haven’t found it? It’s estimated that there are more than \(10^{60}\) possible small molecules in chemical space alone, a vast region of chemical space that eclipses anything we can synthesize and screen using the traditional approach to chemistry and materials discovery.
Machine learning for scientific discovery is challenging this paradigm. Machine learning algorithms are capable of predicting the behavior of molecules and sifting through candidate materials and frontiers before scientists spend valuable time in the lab working with them.
In this article, we look at how machine learning algorithms work, where they have had a tangible effect on research in materials science and chemistry, and the necessity of experiments and data in converting a prediction into a discovery.
The Core Mechanics of How ML Intersects with Chemistry and Material Physics
When predicting properties with ML, one of the first things that a model needs to be able to do is to take structure and represent it in a language that the model can understand. Some examples are: a series of words called SMILES for molecules, a series of numbers representing the charges on atoms and distances called the Coulomb matrix, graph network models with atoms as nodes and bonds as edges, 3D voxel models which capture the shape of a molecule.
The next challenge is the cost of quantum calculations. Models such as SchNet, PaiNN and M3GNet learn from existing calculations to estimate properties such as energy, forces and bandgaps, reducing the need to repeat expensive simulations. Microsoft’s Skala 1.1 offers a related but distinct approach. Its deep-learning exchange-correlation functional ranked first in 32 of 55 GMTKN55 benchmark categories, with a weighted average error of 2.8 kcal/mol. It improves DFT, rather than replacing it.
Researchers can change how they search. Forward screening generates candidates and tests their properties. In inverse design, one begins from a desired property and designs structures which might be able to exhibit it, typically via generative models. For small data sets, techniques like Random Forests and Gaussian Processes can be reasonable; large data sets could work well with deep equivariant neural networks.
Real-World Applications Transforming Industry Frontiers
The true potential of AI materials discovery becomes obvious when research shifts from simulation on the computer and turns to solving real industry problems. Energy storage technologies can be an example here. It is possible to utilize machine learning algorithms to examine solid electrolytes with regard to ion conductivity, learn about solid electrolyte interface formation, and discover materials capable of inhibiting dendrite formation. In addition, it is useful when searching for cathodes that do not require rare metals. Nevertheless, each promising candidate should be tested electrochemically to prove its worth.
Another example where machine learning might be used is catalysis. Machine learning models can simulate molecular interactions with the surface of the catalyst to find materials that will boost hydrogen production through hydrogen evolution reaction. The same applies to finding materials that can catch or convert carbon dioxide into something useful.
Materials used in electronics offer a third frontier. High-thermal-conductivity polymers, dielectric materials and organic semiconductors all require specific combinations of properties. Microsoft’s MatterSim-based workflow demonstrates how this search can narrow down candidates. It screened more than 240,000 materials and identified tantalum phosphorus (TaP) as a potential high-thermal-conductivity material. Experimental measurements recorded its thermal conductivity at 152 W/m/K. TaP was already known, but this research identified a promising property that had not previously received attention. This is where AI for scientific discovery becomes useful, connecting computational screening with experimentally measurable material performance.
The Closed Loop of Integrating ML with Autonomous Self-Driving Laboratories
A promising prediction still has to survive the laboratory. This is where autonomous labs could reinvent the research process. There are four stages into a seamless loop in their vision. Generative AI creates candidate materials, robot systems conduct the synthesis, automated spectroscopy and X-ray diffraction tests their results, and Bayesian active learning reviews the data for the next test.
The benefit is that each experiment can inform the next. Rather than conducting a broad, random experiment, active learning can determine which experiment will offer the most information and erase as much uncertainty as possible. That ensures that scientists aren’t throwing away reagents or hours of lab time on candidates that will add little, if any, value. This also tackles one of virtual screening’s biggest weaknesses: a material’s apparent stability in a simulation that could easily not crystallize or synthesize in real-world lab conditions.
One example of how that decision-making could be done can be seen in IBM’s work. The firm used first-principles electron-transport calculations, semi-supervised crystal graph neural networks and Bayesian optimization to identify candidate topological materials for semiconductor back-end-of-line interconnects at dimensions below 10 nanometres. Such techniques help scale down the universe of possibilities before physical testing starts. The greater potential, perhaps, is to link those predictions with automated experiments so the systems learns from real-world results.
Technical Bottlenecks and Scientific Guardrails
All machine learning models are only as good as the conditions and data they are able to interpret. This poses challenges for researchers when they seek to understand new chemical structures and composition of materials. Existing models are likely to make poor predictions for new and out of distribution cases, occasionally estimating a new structure to be mathematically promising that is then physically impossible to produce.
Data quality creates another challenge. Experimental datasets of higher fidelity continue to be a challenge to capture, costly to create and often siloed in academic repositories. Published papers also tend to show only successful syntheses, underreporting failed attempts. This limits the information models receive about what does not work.
There is also a gap between theoretical stability and practical synthesizability. A structure may appear thermodynamically stable in a simulation, yet the reaction conditions may prevent researchers from producing it. Microsoft’s RetroChimera addresses part of this challenge by predicting plausible synthesis routes. It achieved 90% expert acceptance in blind evaluations, where chemists assessed the proposed routes. However, this does not represent a 90% laboratory synthesis success rate. Ultimately, AI predictions need experimental proof before researchers can treat them as scientific discoveries.
Strategic Implementation for Research Organizations
Research organizations need to create quality data bases before applying scalable AI to scientific problems. FAIR principles make data sets more discoverable, accessible, interoperable, reusable, as well as transparent about the experiments that generated the data. Of course, there should be comprehensive records of results, showing how the underlying evidence was obtained.
IBM considers these data principles when designing a four-step data pipeline for chemistry and materials research that introduces candidate data sets, assesses the scientific quality, establishes expert validation and generates new data where existing evidence isn’t enough.
Also Read: Guidewire to Acquire UK Insurance AI Firm hyperexponential
Moreover, researchers may be able to train physics-informed neural networks (PINNs) that consider conservation laws, thermodynamic properties, and symmetries in order to create models that are more physically consistent and realistic. This would require cooperation among chemists, materials scientists, and machine learning experts.
The Future of Scientific Discovery Lies in Human-AI Collaboration
AI will not instantly bring predictable research to the discovery of new medicines or new materials. Models can guide exploration, highlight potential new structures, and cut down on the number of repetitive experiments, but they can’t decide if this will be useful, or marketable. That’s for scientists who grasp the physics and chemistry that underpins it. The ability to free up researchers’ time, so they can challenge prediction, investigate anomalies, and try out ideas is the true promise of AI. It can take on much of the brute-force searching, but expertise and intuition will still be needed.
English