Interpretable physics-informed retrieval-augmented generation language model for end-to-end inorganic crystal synthesis planning

For decades, the discovery of new solid materials has been a game of two halves. On one side, powerful computers can predict the existence of countless new crystals, calculating how their atoms might arrange themselves to form stable structures. On the other side, human chemists must figure out how to actually build them in a lab. This second step is notoriously difficult. Knowing that a material could exist is very different from knowing how to make it. A computer might tell you a new crystal is stable, but it cannot tell you which powders to mix, what temperature to heat them to, or whether the process requires a special vacuum or a crushing pressure. For a long time, these two worlds—the theoretical prediction and the practical recipe—have evolved separately, leaving a vast gap between what we can imagine and what we can create.

A team of researchers has now built a bridge across this gap using a new kind of artificial intelligence. They developed a system that does not just guess; it reads the history of chemistry to find the right path forward. Instead of treating the creation of a new material as a simple yes-or-no question, the system treats it as a complex planning problem. It looks at a target material and asks: "What have we made before that looks like this?" It then pulls up specific, real-world recipes from thousands of past experiments. By combining these historical records with the laws of physics, the system can suggest a step-by-step plan for making a new crystal, including the exact ingredients and conditions needed.

The core of this work is a massive, organized library of knowledge called the Structured Synthesis Knowledge Base. The researchers fed this system information on 13,820 different inorganic crystals that have already been successfully made in laboratories around the world. They did not just store the names of these materials; they broke down the stories of how they were made. For each crystal, the system recorded the specific method used, the starting chemicals, the temperatures, and the pressures. To make this data useful for a computer, they translated the complex language of crystal structures into plain descriptions that the AI could understand, linking the shape of the atoms to the methods used to assemble them.

When a scientist wants to create a new material, this system, which they call PIRAG-LM, begins by looking for precedents. It does not simply search for words that look similar. Instead, it uses three different lenses to find the best matches. First, it checks chemical similarity, looking for materials made of the same types of atoms. Second, it checks structural similarity, finding materials that share the same geometric arrangement of atoms. Third, and perhaps most importantly, it checks thermodynamic similarity, which relates to how much energy the material holds and how stable it is. By combining these three perspectives, the system finds the most relevant historical examples, even if they were made decades ago or in a completely different lab.

Once the system finds these historical precedents, it uses a large language model to act as a reasoning engine. It takes the recipes from the past and adapts them to the new material. It proposes a list of candidate routes, suggesting which chemicals to mix and what conditions to apply. Crucially, it does not just give an answer; it explains its reasoning. It can point to a specific past experiment and say, "We suggest this temperature because a similar material was made at this heat," or "We recommend this chemical because it worked for a material with the same atomic structure." This makes the process transparent, allowing human experts to see exactly where the advice comes from and verify it.

The researchers tested this system by asking it to predict how to make materials that had been reported in scientific literature between 2021 and 2026, using only knowledge available before 2021. The system was remarkably accurate. It correctly predicted the general method of synthesis for 87.2% of the materials, a significant improvement over using a language model alone, which only got 72.1% right. When the system was further refined to include specific data about high-pressure stability and thin-film growth techniques, its accuracy rose to 91.4%. This proved that the system was not just memorizing text but was truly understanding the physical relationships between materials and the methods used to create them.

To prove that this approach works in the real world, the team did not stop at computer simulations. They took five completely new compounds that had never been made before and asked the system to design a synthesis plan for each. These materials included potential fuel cell components and candidates for controlling how objects expand or contract with heat. The system generated detailed plans, suggesting specific starting powders and heating schedules. The researchers then went into the lab and followed these plans, refining the conditions based on expert assessment and experimental feedback. They successfully created all five new compounds. The crystals they produced matched the theoretical predictions, confirming that the AI had provided a viable roadmap from theory to reality, while expert iteration was necessary to identify the specific workable conditions for each material.

What makes this achievement particularly powerful is how the system handles its own limitations. If the system encounters a material that requires a very high pressure or a special thin-film technique that is not well-represented in its library, it can be updated immediately. The researchers simply added new records about high-pressure experiments or thin-film growth to the database, and the system instantly learned to suggest those methods for future targets. There was no need to retrain the entire AI or start over. This flexibility means the tool can grow and improve as scientists discover more about the physical world, constantly narrowing the distance between what we can calculate and what we can build.

The success of this project suggests a new way forward for materials science. It moves the field away from a binary view where a material is either possible or impossible, toward a more nuanced understanding of how to make it. By linking the abstract laws of physics with the messy, practical reality of laboratory experiments, this system provides a clear, evidence-based path for discovery. It shows that when artificial intelligence is grounded in physical reality and historical data, it can do more than just predict; it can guide human hands in the creation of the materials of the future.