Why AI-Discovered Materials Are Hard to Manufacture at Scale
AI-discovered materials often fail to move from the computer to the factory because the models that predict their properties ignore the practical constraints of synthesis, processing, and cost. A machine learning model can propose a crystal structure with excellent theoretical strength or conductivity, but that structure may require extreme pressures, rare elements, or a synthesis route that does not exist yet. This article explains the main reasons behind this scale-up gap, what researchers are doing about it, and what it means for the future of materials science.
How AI Discovers New Materials
AI models for materials discovery are trained on large databases of known compounds, such as the Materials Project or the Open Quantum Materials Database. These databases contain calculated properties—like formation energy, band gap, and elastic modulus—for hundreds of thousands of hypothetical and real crystals. The AI learns patterns that correlate atomic composition and structure with these properties, then generates new candidates by rearranging atoms or combining elements in ways that humans might not try. For example, generative models can propose thousands of new ternary or quaternary compounds in a single run, each with a predicted property of interest.
However, the training data are almost entirely computed properties, not measured ones. The calculations assume perfect crystals at zero temperature and ignore defects, impurities, and the actual conditions of a chemical reactor. So the AI is essentially optimizing for a theoretical ideal that may not be reachable in practice.
The Synthesis Gap: What the Model Doesn’t Know
The most common reason an AI-discovered material fails at scale is that no one knows how to make it. A predicted crystal structure might be thermodynamically stable at 0 K, but the synthesis route—the sequence of chemical reactions and processing steps—is not part of the model’s output. For instance, a material might require a high-pressure synthesis step (e.g., 10 GPa) that is only possible in a diamond anvil cell, which produces milligrams at best. Scaling that to kilograms is currently impossible.
Even when a synthesis route exists, the AI does not predict the yield, purity, or phase stability under real-world conditions. A compound that is predicted to be stable might decompose at the temperatures needed to sinter it into a solid part, or it might react with the crucible, or it might form a different polymorph when cooled slowly. These are all issues that require experimental iteration, which is slow and expensive.
Processing and Manufacturing Constraints
Once a material is synthesized as a powder, it must be processed into a usable form—a film, a fiber, a bulk part, or a coating. This step introduces additional constraints that AI models rarely consider:
- Melting point: If a material melts above 3000°C, conventional furnaces cannot handle it, and alternative methods like spark plasma sintering or arc melting may be needed.
- Mechanical workability: A material that is predicted to be superhard may also be extremely brittle, making it impossible to machine or roll into sheets.
- Thermal stability: The material might be stable at room temperature but decompose or oxidize at the temperatures required for processing (e.g., 1000°C for sintering).
- Chemical compatibility: The material might react with the mold, the atmosphere, or the binder used in additive manufacturing.
These are not exotic concerns; they are the everyday realities of manufacturing. A model that only predicts a single property—like a high superconducting transition temperature—ignores all of these, so the material may be scientifically interesting but industrially useless.
Cost and Scalability of Raw Materials
AI models often propose materials containing rare or expensive elements because those elements have interesting electronic configurations. For example, a high-temperature superconductor might include yttrium, barium, copper, and oxygen (YBCO), which is already known and can be made, but scaling it to long wires is difficult because of the brittle ceramic nature. Other proposals might use ruthenium, iridium, or scandium—elements that are scarce and cost thousands of dollars per kilogram. Even if the material can be synthesized in a lab, the cost of the raw materials alone can make commercial production impossible.
In addition, the synthesis process itself may require expensive reagents, solvents, or energy inputs. For example, a material that needs to be grown as a single crystal from a flux at 1500°C for two weeks will never be cheap. AI models that do not include cost as a variable will happily propose such materials, but no company will invest in scaling them up.
Characterization and Quality Control Difficulties
Even after a material is synthesized and processed, it must be characterized to confirm that it has the predicted structure and properties. This is not trivial. X-ray diffraction can confirm the crystal structure, but the actual material may contain defects, grain boundaries, or secondary phases that degrade its performance. For example, a predicted high-capacity battery cathode might have a theoretical capacity of 300 mAh/g, but in practice, the real material might only reach 200 mAh/g because of structural degradation during cycling.
Quality control at scale is even harder. In a lab, you can carefully control every parameter and check each sample. In a factory, you are making tons of material, and variations in temperature, pressure, and raw material purity will lead to batch-to-batch variability. This is a major barrier for any new material, but it is especially acute for AI-discovered ones because there is no prior manufacturing experience to draw on.
Bridging the Gap: New Approaches
Researchers are aware of these limitations and are developing new methods to make AI predictions more manufacturable. One approach is to include synthesis-related features in the training data, such as the temperature and pressure at which a compound was made, or the synthesis method used. Another is to use active learning, where the AI proposes a material, a robot synthesizes it, and the result is fed back into the model to improve its predictions. Some groups are also using AI to optimize processing parameters, such as sintering temperature or cooling rate, to achieve the desired microstructure.
However, these methods are still in their infancy. The fundamental problem is that materials science is a multi-scale, multi-physics problem: the AI predicts the atomic-scale structure, but the manufacturing process involves fluid dynamics, heat transfer, and chemical kinetics at the macro-scale. No single model can capture all of that, so the gap is likely to persist for some time.
Frequently Asked Questions
Why can’t AI just figure out how to make the material?
AI can be trained to predict synthesis routes, but the data on successful syntheses are sparse and inconsistent. Most published papers do not report failed attempts, so the model does not know what does not work. Also, synthesis is a sequential decision process with many variables, and current AI models are not good at planning multi-step actions in the physical world.
Are there any AI-discovered materials that have been scaled up?
Yes, a few. For example, a new lithium superionic conductor called Li6PS5Cl (a lithium argyrodite) was identified with the help of high-throughput screening and is now being considered for solid-state batteries. However, it took years of additional experimental work to optimize the synthesis and processing, and it is still not in commercial production.
Is the problem that AI is not smart enough?
No, the problem is that the training data are incomplete and the models are not designed to handle the full complexity of manufacturing. AI is very good at finding patterns in data, but it cannot extrapolate to conditions that are not in the training set. Manufacturing involves many such conditions.
Could we just use AI to design the manufacturing process too?
In principle, yes, but the data needed to train such a model are even scarcer. Manufacturing process data are often proprietary and not shared, and the number of variables is enormous. It is an active research area, but not yet practical.
Will this gap ever be closed?
Probably, but it will require close collaboration between AI researchers, materials scientists, and chemical engineers. It will also require the creation of large, open databases of synthesis and processing data, which is a major undertaking. The gap is not fundamental; it is a matter of data and effort.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.