In our latest episode of Lexicon, we sat down with Joshua Young, a senior application scientist at Matlantis, to explore how artificial intelligence (AI) is reshaping the way new materials are discovered. Materials science has always been a foundation of modern technology, but today it is increasingly driven by simulation, data, and machine learning.
Young explained how AI-accelerated simulations are helping researchers screen millions of potential materials at unprecedented speed, why many teams are willing to trade perfect accuracy for faster results, and why fully autonomous discovery remains out of reach.
He also shared why trust, intellectual property protection, and human expertise remain central as AI becomes embedded in real-world materials research.
Also, subscribe to IE+ for premium insights and exclusive content!
AI as an accelerator
When people hear the phrase AI in materials science, Young explained, they often imagine “end-to-end complete discovery of materials.” In practice, however, AI tools are being layered onto existing workflows rather than replacing them wholesale.
“Our survey right now shows that many researchers are blending these emerging AI simulation tools with the more traditional physics-based methods like density functional theory,” he explained.
Young added that about half of the teams surveyed are already using AI-native simulation platforms in production, even as traditional methods still account for a slight majority of total workloads.
“This kind of indicates that AI is currently coexisting with and augmenting established techniques,” Young said, “mainly to speed up specific parts of research, more than kind of completely taking over the discovery process from end to end.”
Why full autonomy still eludes us
While AI-guided simulations and experiments are technically possible today, Young stresses that they remain fragmented. “The major hurdle is right now, they’re still kind of segmented and compartmentalized,” he explains.
Simulations produce text-based outputs. Experiments generate images, spectra, and microscopy data. Integrating these disparate data streams into a single closed-loop system is extremely challenging. “While all of these things exist, linking them all together is where the real challenge lies in the autonomous discovery workspace,” he added.
As a result, Young estimates that widespread, fully autonomous materials discovery across broad domains is still “probably still a few years away, maybe three to five years away.”
Humans in the loop
Even as automation advances, Young is confident that human expertise remains essential. “We kind of call this human in the loop, right?” he said. “You do not need the human intuition. We need the domain expertise still to interpret and run experiments,” he added.
While AI agents are increasingly being used to automate parts of simulation and analysis, Young argues that removing humans entirely would likely lead to poor outcomes.
“Fully autonomous is perhaps the goal,” he said, “but at some point still, I think human in the loop is the way to go to achieve actual speed up and materials discovery,” he added.
One key reason, he explained, is trust. According to the survey, only about 14% of researchers report very high confidence in AI systems. “Some of this hesitation stems from a lack of very confident trust,” Young explains. Without expert oversight, there’s a real fear of “junk in, junk out,” he added.
Speed over perfection
One of the most striking findings from the Matlantis survey is how willing researchers are to sacrifice precision in exchange for speed. “73% of researchers would trade a bit of accuracy for 100× faster results,” Young noted.
For computational specialists, this typically means accepting deviations of five to ten millielectronvolts per atom, which area small enough area to preserve meaningful trends. “The willingness to accept these deviations reveals that the industry is intensely focused on throughput and screening capacity over absolute perfection,” he told us.
What’s more, the pressure is real. “94% of teams surveyed had to abandon some simulation project due to time constraints,” Young said. As a result, speed is no longer just a convenience; it’s a prerequisite for innovation.
“It signals a pragmatic shift,” he explained, “where speed is viewed as the primary enabler for unlocking innovation that is currently stuck in computational bottlenecks.”
Young also pushed back against the idea that simulations must deliver perfect predictions to be useful. Citing an interview with Professor Yukihiro Shimogaki of the University of Tokyo, he notes that “accuracy shouldn’t really be about hitting like an exact number.”
Instead, simulations are most valuable when they “correctly predict qualitative trends and narrow down the search space.” In this framing, AI-powered simulations act as filters, rapidly eliminating weak candidates so that experimental resources can be focused where they matter most.
Real-world wins
Beyond theory, AI-accelerated simulation is already delivering tangible benefits. “These AI-enhanced simulations are already delivering measurable value,” Young told us. Organizations using them report average savings of around $109,000 per project, driven by reduced physical experiments, lower compute costs, and faster iteration.
In one case study, a chemical company evaluated 13 potential catalyst improvements. “With conventional methods, it would have taken them two or three years,” Young said.
Using AI-driven simulations, they reached the same conclusion in just six weeks. “They were actually happy,” he added, because they avoided years of wasted effort.
In another example, researchers screened 32 million potential solid-state battery materials in under a week. “Traditionally, this would have taken decades, maybe,” Young added. The resulting prototype battery was developed in under nine months.
Trust, intellectual property, and security
Despite these gains, trust remains a gating factor, particularly around intellectual property. “Data security is basically non-negotiable,” Young said. Not a single respondent in the survey viewed IP protection as a non-issue.
“These materials breakthroughs can be worth billions of dollars,” he explained, which makes fears of data leakage or model exposure entirely rational. As a result, companies increasingly demand strict safeguards, private cloud deployments, or on-premise solutions.
Young emphasizes that customers retain full ownership of their data. “We don’t involve ourselves with any of the IP-related matters. All IP remains with the user,” he added.
The road ahead
Looking five years into the future, Young believes that massive, high-speed screening will become routine. “Running thousands, hundreds of thousands, millions of simulations in days or weeks,” he said, “will probably become the standard baseline for initiating any new projects.”
He also envisions the rise of self-driving laboratories, or integrated systems where simulations trigger robotic experiments, and experimental results feed directly back into AI models. “Turning even the physical failures into assets,” as he put it.
Crucially, Young does not see this future as one that sidelines scientists. “Rather than replacing the scientists, this will really free the scientists and researchers to pursue novel ideas,” he said.
By automating the laborious parts of research, AI allows humans to focus on what they do best. “All of this laborious time can be turned into innovation,” Young concluded. “That’s what we humans are good at: coming up with plans, strategies, and novel ideas.”