Learning the Chronology of Materials Discovery for Synthesizability Prediction

The search for new materials is a race against time. Scientists can now use powerful computers to imagine millions of new chemical recipes, proposing candidates for batteries, solar panels, and medicines far faster than any laboratory can test them. This creates a bottleneck: a vast ocean of theoretical possibilities where only a tiny fraction can ever be built and proven to work. To navigate this, researchers have developed computer models that try to guess which of these imaginary recipes are actually realizable in a lab. Traditionally, these models treat the history of science as a static list, looking only at what has been made and what has not, without considering when or why discoveries happened. They assume that the path to a new material is random, ignoring the fact that human discovery follows a specific, evolving story shaped by available tools, funding, and shifting scientific interests.

A team of researchers at Imperial College London has introduced a new approach called ChronoSyn that changes how these predictions are made. Instead of treating the history of materials as a simple collection of facts, this model learns from the timeline of discovery itself. It recognizes that the order in which scientists found new materials is not random; it reflects a progression driven by practical needs and technological advances. By teaching the computer to understand this historical sequence, the model gains a deeper insight into which chemical combinations are likely to be synthesized next. The researchers found that by incorporating the dates of past discoveries, the model became significantly better at identifying promising new materials, even those that had never been seen before.

The core of this work lies in how the model processes information. It takes a chemical composition—a specific mix of elements—and converts it into a digital fingerprint. This fingerprint is then fed into two different parts of the computer program simultaneously. One part tries to predict whether the material can be synthesized, while the other part tries to guess when that material was likely discovered relative to others. The model learns by comparing pairs of known materials, adjusting its internal logic so that it assigns a "later" score to materials reported in recent years and an "earlier" score to those found long ago. This dual task forces the computer to build a shared understanding of chemistry that respects the flow of time. The result is a system that does not just memorize a list of known materials but understands the trajectory of how science moves through chemical space.

When tested against standard models that ignore the timeline, ChronoSyn proved to be more accurate. In experiments using databases containing hundreds of thousands of known and hypothetical materials, the new model correctly identified a higher percentage of real, synthesizable compounds. The improvement was not just a small statistical bump; it was a clear signal that the history of discovery provides valuable clues. The researchers observed that the model learned to group materials in a way that mirrored their actual history. In the computer's internal map, materials discovered in the 1970s and 1980s clustered in one area, while those found in the 2010s and 2020s gathered in another. This organization allowed the model to distinguish between chemical families that have been thoroughly explored and those that are just beginning to be investigated.

This ability to read the timeline proved particularly useful when looking at materials that had not yet been reported. The model identified groups of hypothetical compositions that resembled the chemical families discovered in recent decades. These "recent-like" materials were given higher scores for synthesizability than older-looking ones. The researchers found that this made sense: chemical spaces that have been actively studied recently, such as those used for lithium-ion batteries or fuel cells, are still being explored, meaning there is a higher chance that unreported materials in these areas are actually realizable. Conversely, chemical spaces that were heavily studied decades ago but have since gone quiet are more likely to contain materials that are genuinely difficult or impossible to make. The model effectively learned to prioritize the "frontier" of current research over the "backwater" of past exploration.

To test the practical value of this approach, the researchers performed a retrospective experiment. They trained the model using only the knowledge available up to a specific year, such as 2018, and then asked it to predict which materials would be discovered after that date. ChronoSyn consistently outperformed the standard models in identifying these future discoveries. It successfully flagged compositions that were indeed reported in the scientific literature in the years following the cutoff, demonstrating that the model could use the past to forecast the future. This suggests that the path scientists have taken to find materials so far holds a map to where the next breakthroughs will occur.

The study also revealed that the model was particularly good at identifying complex materials containing rare elements, such as those with f-block elements like lanthanides. These materials often appear in distinct waves of discovery, with intense periods of research followed by quieter times. Because the model learned the timing of these waves, it could better recognize when a new composition fit into a currently active research trend. When the team applied ChronoSyn to a massive hypothetical space of 50 million potential chemical combinations, it highlighted specific families of compounds that were likely to be realizable. A check against scientific records confirmed that many of these high-scoring predictions were indeed materials that had been synthesized and reported in the past but were missing from the main databases used for training.

The researchers acknowledge that their method has limits. The model relies on publication dates, which are a proxy for discovery and may not always match the exact moment a material was first made. It also reflects the biases of human research, potentially favoring chemical areas that have historically attracted attention while overlooking promising but neglected fields. Furthermore, the model predicts whether a chemical formula is possible, but it does not yet account for the specific crystal structures that might make a material stable or unstable. Despite these constraints, the work demonstrates that the history of science is not just a record of what has been done, but a guide for what can be done next. By learning the chronology of materials discovery, scientists can now use the past to more effectively navigate the vast and uncharted territory of the future.