Can machines discover?

How the Materials Research Institute is both using and enhancing AI in materials research

By Jamie Oberdick  

For decades, materials scientists have relied on knowledge, experience, and sometimes painstaking trial and error to understand how matter behaves. Progress often came slowly, driven by incremental advances and hard-won insights. Today, artificial intelligence is accelerating that process, revealing hidden relationships within complex data and opening new pathways for discovery. For some researchers at the Materials Research Institute (MRI), AI and machine learning are no longer experimental add-ons. They are becoming integral collaborators in the scientific process.   

At the same time, the rise of AI has introduced new challenges. As algorithms grow more powerful, they also demand more energy, more computational resources, and more physical infrastructure. At Penn State, researchers are grappling with both sides of this transformation, using AI to push the boundaries of materials science while also rethinking how computing itself should be designed for a more sustainable future.   

From intuition to algorithms

Stephanie Law, associate professor of materials science and engineering and Wilson Faculty Fellow, develops custom semiconductor and quantum materials with carefully engineered electronic and optical properties, using atomic-scale growth techniques to push the limits of sensors, electronics, and photonics.   

That level of control comes with extraordinary complexity. Even small changes in growth conditions can dramatically alter a material’s structure and performance, making it difficult to predict outcomes using traditional approaches.   

“One of our major challenges is finding the optimal synthesis parameters for a given material,” Law said. “There are multiple parameters to control, and each parameter has a wide range of values, making parameter space extremely large. Traditional material optimization followed a trial-and-error approach, where we would try a set of parameters, measure the material properties, adjust the parameters, measure again, and repeat. This process was slow, and we were not guaranteed to find the optimal set of parameters. Because the synthesis parameters are interdependent, it was also difficult to isolate the effect of one variable on a particular material property.”   

Rather than relying solely on trial and error, Law’s group has begun turning to machine learning to navigate that complexity. AI models can evaluate many variables simultaneously, uncovering relationships that would be nearly impossible to identify by hand. 

“We have been collaborating with Wes Reinhart (assistant professor of materials science and engineering and Institute for Computational and Data Sciences faculty co-hire) to address both problems,” she said. “We are using Bayesian optimization, a type of machine learning, to help us intelligently explore and understand parameter space. The model suggests initial parameter sets to try, then uses the data on material properties to suggest optimal growth parameters.”   For Law, the implications extend beyond any single experiment. She sees machine learning reshaping how materials research is conducted at a fundamental level.   “I think machine learning will become integrated into workflows that relate synthesis conditions with material structure and properties,” she said. “These experiments are expensive and time-consuming, making it very appealing to use machine learning to optimize synthesis and to understand relationships among parameters. It will enable us to synthesize more materials with tailored properties, which can then be used to create a wide range of new devices.”   

Seeing the invisible   

While some researchers focus on creating new materials, others are tasked with understanding these materials at a deep molecular level once they exist. For Danielle Reifsnyder Hickey, assistant professor of chemistry and materials science and engineering, that challenge unfolds at the atomic scale, where crucial structural details are often hidden within noisy experimental data.

“My research group focuses on understanding materials’ structures and properties, primarily for electronics and energy applications,” Hickey said. “We aim to develop an atomic-scale understanding of how materials work, which is important for designing new and better technologies. 

For instance, controlling atomic structures and developing new materials can determine how fast our electronics are, how much they cost, and whether they are small and light enough to carry around easily.”   

Extracting that information is rarely straightforward. Experimental measurements can be incomplete or ambiguous, especially when materials are disordered or undergoing rapid transformations. Machine learning has become a powerful tool for cutting through that uncertainty. Hickey’s group also uses AI to study how materials evolve over time, capturing transitions that would otherwise be difficult to follow.   

“We use machine learning to precisely determine the atomic positions in materials,” she said. “Machine learning can help us identify a material’s crystal structure, shape, and anomalies such as defects. These anomalies are often key to understanding how materials behave.” 

Even so, Hickey emphasizes that AI and machine learning are not a substitute for scientific judgment. The same tools that reveal new insights can also obscure physical meaning if their limitations are not well understood. 

“Machine learning is great at recognizing patterns and distinguishing subtle differences between related signals, which can be challenging to identify by eye,” Hickey said. “These small variations in materials’ structures can control their properties, so it is important to understand them precisely.”  

Hickey added, “AI is permeating many aspects of science, and this will accelerate as AI tools keep improving. By bringing together huge amounts of information quickly, it will keep accelerating the pace at which we can make new connections between information and discover new physical phenomena.”   

Algorithms as instruments

In Wes Reinhart’s view, who Law mentioned earlier, AI represents a shift in how scientists interact with data. Rather than functioning as a black box, AI serves as a new kind of scientific instrument, one capable of probing complex relationships that traditional methods struggle to capture. This is part of what drives Reinhart’s research into AI as a science tool. 

“Real-world data is messy and limited,” Reinhart said. “Standard ML models built for extremely abundant image and text data available on the internet don’t work the same way on the small, highly specialized data we have in materials science. Some of that data comes from special instruments that only exist in one or two labs in the world. My research group is developing frameworks specifically designed to handle these messy realities and make AI useful to synthesis scientists.” 

In Reinhart’s view, the goal is not to replace human intuition but to augment it. AI can help researchers identify promising directions and ask better questions, accelerating the pace of discovery. 

“Domain experts need to trust what the AI is telling them before they make experimental decisions based on those predictions. Our focus is on developing explainable AI that lets scientists see why the model is making certain predictions.”

“We’re pursuing two main strategies,” Reinhart said. “First, we use large language models to help design materials and experiments. These models have been trained on enormous amounts of scientific text, which allows them to suggest new molecules or material modifications that a human researcher might not immediately consider. We can then validate those ideas using simulations.” 

In practice, Reinhart explained, the real power of these models is their ability to respond to plain-language prompts. Instead of specifying every parameter mathematically, researchers can describe the properties they want, and the model can propose changes that move the material in that direction. His group has also found that large language models can outperform other machine-learning approaches when it comes to planning efficient sequences of experiments to collect new data. 

“The second area is using large language models to parse the scientific literature,” Reinhart said. “We’re encoding information from hundreds or even thousands of published papers to extract knowledge that no single person could realistically read and synthesize. In recent work, we found that representing metal processing conditions using natural language actually improves predictive performance, saving time and accelerating the discovery of high-performance materials.”

Reinhart notes that the primary benefit of this work is accelerating materials discovery while reducing costs. Tools like visual “what-if” models allow scientists to test hypotheses virtually before running experiments, while large language model-based approaches help AI systems parse vast amounts of scientific literature. Multimodal models can also reconstruct and connect different types of data, such as linking spectral information with micrographs. Together, these capabilities act as force multipliers, allowing experimental scientists to work more efficiently and focus their time on the most promising directions.

Looking ahead, AI is unlikely to replace traditional laboratories but will instead serve as an essential tool within them.

“This will require a cultural shift in our workforce,” Reinhart said. “A major goal of my work is to help train scientists who are ‘bilingual.’ That is, they need to be equally fluent in materials science and AI/ML. I think both the physical sciences and data sciences will benefit from new ideas in data processing and learning algorithms developed by these transdisciplinary researchers working on materials science’s grand challenges.”   

The brain as blueprint   

Saptarshi Das, professor of engineering science and mechanics and materials science and engineering, sees the rapid adoption of AI as a fundamental shift rather than a gradual evolution. The speed at which algorithms are transforming research has reshaped expectations across the field. 

“Things are moving very, very fast,” Das said. “In our own research activities, we are already experiencing what AI can do. It’s not just about automation anymore. It’s about fundamentally changing how we think about discovery.”  

Much of Das’s work focuses on neuromorphic computing, an approach inspired by the human brain. Compared to conventional digital computers, the brain performs complex tasks while consuming remarkably little energy.

“If you look at the brain, it’s extraordinarily efficient,” he said. “It performs complex computations using a fraction of the energy consumed by today’s digital computers. One of the big questions is whether we can design materials and devices that operate in a similarly efficient way.” 

At the same time, Das is deeply aware of the broader implications of AI’s growth. As computing demands increase, questions of sustainability, equity, and long-term impact become impossible to separate from technical performance. 

“The question is not just what AI can do,” he said. “The question is how we sustain it. How do we build systems that are powerful but also responsible, equitable, and environmentally sustainable? That’s a much bigger challenge than any single technological breakthrough.” 

For Das, materials research sits at the heart of that challenge.

 “If we want AI to be sustainable, we can’t just rely on better software,” he said. “We need fundamentally new materials, new device architectures and new ways of thinking about computation. That’s where places like the Materials Research Institute come in. We’re not just using AI as a tool. We’re helping to define what the next generation of AI will physically look like.”

Beyond the hype

At the Materials Research Institute, the question is no longer whether machines can assist discovery, but how deeply they can be integrated into the scientific process. Across disciplines, AI is reshaping how materials are designed, analyzed, and understood, not by replacing human insight, but by expanding it. When algorithms help navigate vast parameter spaces, reveal hidden atomic structures, or suggest experiments in plain language, discovery becomes faster, more targeted, and more collaborative than ever before.

At the same time, MRI researchers are confronting the growing challenges that accompany AI’s rise, from escalating power consumption and energy-intensive data centers to questions of transparency, trust, and long-term sustainability. The future of AI-driven science will depend not only on smarter algorithms, but on new materials, device architectures, and computing paradigms that make intelligence itself more efficient and responsible. By pairing technical innovation with critical reflection, MRI is helping shape a future where discovery emerges from a thoughtful partnership between intelligent humans and intelligent machines designed with physical, environmental, and societal limits in mind.