Combining physical models with dynamically acquired experimental information for the optimization of multicomponent NASICON fast ionic conductors in a self-driving laboratory

The quest for better batteries often leads scientists to look at the materials that let electricity flow inside them. In a sodium-ion battery, which offers a sustainable alternative to the lithium-ion batteries powering our phones and cars, this flow happens through a solid electrolyte. This solid material must be a highway for sodium ions, allowing them to zip through quickly to charge and discharge the battery, while remaining stable and easy to manufacture. One promising family of materials for this job is called NASICON. These are complex crystals made of sodium, phosphorus, oxygen, and various metal atoms. The challenge lies in their complexity: by swapping different metal atoms into the crystal structure, scientists can tune how fast the ions move. However, finding the perfect mix is like searching for a needle in a haystack that keeps changing shape. If the mix is slightly off, the material refuses to form the desired crystal structure and instead creates a jumble of useless impurities, rendering it useless for a battery.

Traditionally, finding these perfect mixes has been a slow process of trial and error, where researchers guess a formula, bake it in a furnace, and hope for the best. If it fails, they guess again. This approach is inefficient when the number of possible combinations is vast. A new study introduces a different way to work. Researchers have built a system that combines what we already know about physics with real-time feedback from experiments. They created a framework that acts like a self-correcting guide. It starts with a computer model that predicts which chemical recipes are likely to work and which are likely to fail. Then, it sends a candidate recipe to a robotic laboratory that mixes the powders and heats them. The robot then analyzes the result using X-rays to see what actually formed. If the result is not perfect, the system learns from the mistake, adjusts its internal map of what works, and immediately designs a better recipe for the next round. This cycle repeats, with the machine learning from every success and every failure to navigate the complex chemical landscape much faster than a human could.

In this study, the team applied this self-driving approach to discover new sodium conductors. They set the robot loose in a high-dimensional chemical space, testing dozens of variations of NASICON materials. The system was designed to balance two competing goals: finding a material that is easy to synthesize without impurities, and finding one that conducts electricity well. The robot ran 78 trials in total. In the first round, it used its initial knowledge to pick starting points. Many of these first attempts resulted in materials that were mostly the right crystal structure but still contained some unwanted impurities. The system did not discard these results; instead, it analyzed the specific impurities that appeared. It then used that information to tweak the chemical recipe for the next round, reducing the ingredients that caused the trouble while keeping the ingredients that helped form the desired crystal.

The results of this automated campaign were significant. After just one or two rounds of adjustment for each chemical system, the robot successfully synthesized 23 compositions where the desired NASICON crystal was the dominant material. Of these, 18 were located in a specific region of the chemical space that the computer models predicted would have high ionic conductivity. To verify these predictions, the researchers took the most promising samples and measured how well they conducted electricity. Two of the new materials stood out. One showed a total ionic conductivity of 0.7 millisiemens per centimeter, and another reached 0.3 millisiemens per centimeter. When the researchers looked closer at the internal structure of the material, ignoring the boundaries between grains, the conductivity of the first sample jumped to 3.96 millisiemens per centimeter, and the second to 3.17 millisiemens per centimeter. These numbers are competitive with the best known sodium conductors, and they were achieved without the years of manual optimization usually required.

The study also revealed something important about the limits of these materials. By observing which recipes worked and which failed, the researchers were able to map out the boundaries of stability for these crystals. They found that the amount of energy a material can have above its most stable state before it fails to form depends on the specific ingredients used. For materials made only of phosphate groups, the limit was quite strict. However, for materials that mixed phosphate with silicate groups, the system could tolerate a much wider range of instability before failing. This suggests that mixing different types of chemical building blocks can open up new pathways for synthesis that were previously thought to be too difficult. The researchers noted that while the system found many successful recipes, it also identified cases where a material looked good on paper but failed to conduct well, highlighting that the relationship between the chemical recipe and the final performance is still complex and not fully understood.

The success of this project demonstrates the power of closing the loop between theory and experiment. Instead of relying on static computer models that might miss subtle real-world behaviors, or relying solely on human intuition, the system used a dynamic approach. It treated every experiment, even the failed ones, as valuable data. When a synthesis failed to produce a pure crystal, the system didn't just move on; it used the specific nature of the failure to refine its understanding of the chemical rules. This allowed it to quickly home in on the narrow window of compositions that were both synthesizable and functional. The researchers emphasized that this method is not a magic solution that solves all problems instantly. There were still cases where the system struggled to predict the exact composition of a crystal when it was mixed with other phases, and some promising recipes still resulted in low conductivity. However, the ability to rapidly identify and discard dead ends, while simultaneously refining the search for the best candidates, marks a significant step forward.

This work lays the foundation for a new kind of materials discovery. The framework used here is modular, meaning it can be easily adapted to test other properties or materials. The researchers suggest that in the future, such systems could be expanded to consider other factors like the cost of raw materials or how well the battery material interacts with other parts of a battery cell. The ultimate goal is to move beyond simple screening, where thousands of materials are tested one by one, toward intelligent laboratories that can reason about their results and design their own next steps. By proving that a robot can learn from its mistakes to discover new, high-performance materials, this study shows that the future of battery development may lie in machines that do not just follow instructions, but actively learn to solve the puzzles of chemistry.