How do we discover new materials?

By Joseph Montoya and Muratahan Aykol

Just like pilots use flight simulators to learn how to fly a plane before getting in a real cockpit, we also use simulators to learn about materials before conducting expensive experiments. As discussed in an earlier post, TRI’s materials team created Computational Autonomy for Materials Discovery (CAMD), an AI-powered search tool that autonomously runs simulations to find new, undiscovered materials.

In the past year, CAMD has generated thousands of new compounds that are stable, meaning they have a good chance of being synthesizable and in turn will accelerate the discovery of new meaningful materials, at least on the computer! In an ideal world, we would typically make these compounds, characterize them, and test their utility in actual devices, such as Lithium-ion batteries, fuel cells, solar cells and so on. But there is still a bottleneck: synthesis planning.

Cooking Up Inorganic Compounds with Synthesis Planning

Some inorganic compounds occur in nature in a useful form, like the salt (NaCl) on your kitchen table or quartz (SiO2), which was central to the electronics industry decades ago. But naturally occurring compounds are the exception rather than the rule. Lithium cobalt oxide (LiCoO2), for example, which is in the batteries of most of our cellphones, is a lab-made compound, like almost any other technological inorganic compounds.

A synthesis plan (think of a recipe, not just the list of ingredients) is required to make these target compounds that would ultimately lead to applications. The word “target” should not make us believe this is a procedure we have control over, however. On the contrary, most of the inorganic synthesis for making new compounds remain exploratory. With both old and new synthesis techniques we can make a variety of compounds. But this exploration often happens without a prediction for what would come out of a given synthesis procedure.

Piro and the Power of Prediction

The main hurdle for inorganic solids, in comparison to organic molecules, is that the physical process for their formation is extremely difficult, if not impossible, to predict or model with confidence. This difficulty does not mean there are no advances for predictive synthesis in inorganic space. A major thrust in the field has been using machine learning to predict the “synthesizability” of novel compounds and using text mining to collect all of the recipes for a given material in one place.

However, there’s still a big opportunity to make recommendations of synthesis routes on a physical basis, rather than only from statistics or text-mining. Our recently open-sourced software, piro, does just that for solid-state synthesis. More specifically, piro compares different sets of solid reactants, or “precursors”, that might result in a particular crystal when mixed and heated to high temperatures, and predicts which reactions are more likely to form the target crystalline compound.

For example, going back to LiCoO2, we can make it by mixing lithium carbonate (Li2CO3) and cobalt oxide (Co3O4) and heating to around 900℃ (quite a bit hotter than most of our household ovens). Alternatively, we can use lithium peroxide (Li2O2) and another Co-source (say CoCO3) to make the same compound. The famous superconductor YBa2Cu3O7+x (often shortened as YBCO) can be made using metal oxides or carbonates. LiCoO2 and YBCO are so widely studied that we have a pretty good idea of which starting materials work best. What’s exciting is that piro reproduces the experimental observations of efficacy of synthesis reactions not only for these well-studied compounds; but also in way more complicated (and much less studied) cases when such standard precursors may not work well (such as RuSr2GdCu2O8 or K2Mo9S11). If interested in technical details on how piro achieves this, check out our recent publication in JACS (or our preprint here). We’ll explain the main ingredients here.

Lowering the Bar to Increase Chances of Success

For solid-state synthesis, the physical process that describes which compound forms and at what rate is called nucleation. Note that nucleation is not just about synthesis; it governs almost any process in nature where a nucleus of a certain structural arrangement (phase) is formed from assemblage of small building blocks (such as atoms or molecules). For example, CO2 bubbles in your soda nucleate at tiny defects on the surface of the glass. Proteins, polymers, metal alloys, pharmaceutical solids and inorganic compounds also form through nucleation.

In this process, small embryos of crystals need to grow above a certain size and overcome an energy barrier. As a hill is much easier to climb than a mountain, a smaller nucleation barrier is easier to overcome than otherwise. For example, CO2 bubbles in your soda are nucleating on defects of your glass because those defects catalyze nucleation by effectively lowering the barrier (often referred as heterogeneous nucleation). Same will happen during almost any nucleation process, including solid-state synthesis, where certain surfaces would serve as a substrate on which nucleation of our target would have a lower barrier.

Can we find reactant sets that can help nucleate the target compound on their surfaces? That’s the idea behind piro.

The lower the barrier, the faster the potential nucleation of our target, making it the more likely outcome of the reaction. To estimate relative nucleation barriers of different synthesis reactions to form a target, piro combines the physics from the never-aging classical nucleation theory (CNT) with (1) materials data easily accessible from high-throughput computational databases like the Materials Project and (2) quantifying structural and chemical “similarities” between compounds, a familiar concept in machine learning for chemistry.

However, just like a parasite takes nutrients from its host, a parasitic reaction can take over the precursors and make an undesired compound, rather than, say LiCoO2. We use data from open materials simulation databases like the Materials Project to estimate the number of possible parasitic reactions that a given set of precursors could have. This data helps us estimate how likely the reaction will result in the product we want.

With the nucleation barrier vs. the number of parasitic reactions, we now have the opportunity to create something that human beings have been using for thousands of years to navigate: a map of synthesis routes that can optimally lead to our target solid.

A map of LiCoO2 synthesis. The labeled reactions are A: Li+CoO2 and B: Li6CoO4+Co+O2. The leftmost “x” has a nucleation barrier higher than A, and therefore likely to happen more slowly. The rightmost “x” will likely be faster, but less selective than A because it has more parasitic reactions, and will be slower than B because it has a higher barrier. The middle “x” might be worth considering, because it’s close to the Pareto line.

So, here’s a map of LiCoO2 syntheses we created with piro! Rather than latitude and longitude, we have the relative nucleation barrier and the number of parasitic reactions. The closer we are to the bottom of the map, the faster the formation of LiCoO2 should be. The further left we are, the better our chances are for only getting LiCoO2. We can also find the combinations of these properties where we should have the optimal tradeoffs between these two properties we want to minimize, which is a typical Pareto optimality problem.

The real power of our map is that it can explain why some syntheses may or may not work. For example, a reaction may be too slow to nucleate. Or another may suffer from a parasitic reaction consuming all of its precursors. Still another reaction may provide a compromise between these two metrics, so maybe it’s worth trying! Now we can make more informed choices about which synthesis reactions to try and increase our chances for success.

This map isn’t the whole story but so far, piro has been largely successful at identifying good reactions from a sea of reactions we know nothing about. Our hope is that tools like piro will initiate a new pathway for predictive synthesis in the community. The open-source nature of piro allows users to extend it to address other kinds of synthesis techniques beyond conventional solid-state. These tools may also motivate synthesis experts to report more failed syntheses, so we can not only suggest better routes, but also find out where predictive tools are struggling, and improve them.

Lastly, our goal is to integrate a tool like ours with a more comprehensive platform for discovering materials so researchers can filter their results based on what compound might be easiest to make. These are all big undertakings, but they’re also all in progress, so stay tuned!