A new report by Matlantis provides a detailed view of how materials scientists and computational chemists in the United States are weaving simulation and AI into their research.
Study: Accelerating Discovery: AI Trends in Materials R&D – How Simulation Leaders Are Balancing Speed, Accuracy & Trust. Image Credit: vectorfusionart/Shutterstock.com
Drawing on responses from 300 professionals, the data reveals researchers now have access to algorithms capable of accelerating discovery, but their infrastructure and workflows have not kept pace. Computational limits are now slowing the discoveries these tools were meant to accelerate.
Here, we explore the key findings of the Accelerating Discovery report, examining the constraints researchers face today, as well as the opportunities that emerging tools and architectures may create. Fumio Horino, VP Marketing at Matlantis, who played a key role in producing the report, tells us more.
Project Abandonment Due to Compute Constraints
In the past year, 94 percent of teams abandoned at least one planned simulation project because it exceeded runtime expectations or compute budgets.
These numbers clearly show that computational capacity is a significant factor limiting research output, regardless of whether teams use on-premise clusters, private cloud systems, or public HPC resources.
Fumio explains that this issue of abandonment is a result of more than simply infrastructure.
This is a clear mismatch, but the issue is more nuanced than simply “not enough compute.” Organizations are already spending significant amounts on simulation (on average, hundreds of thousands of dollars per year – and most see a strong ROI.
Fumio Horino, VP Marketing at Matlantis
About 52 % of teams halted between one and five projects, while 41 % abandoned between six and ten.
Only around six percent reported no project losses.
These figures are not insignificant – in fact, they paint a stark picture for an industry that relies on fast turnaround times.
A combination of runtime limits, compute budgets, scheduling bottlenecks, data fragmentation, and software or licensing constraints resulted in nearly all teams reporting project abandonment.
This suggests that the industry’s ambitions have grown faster than the supporting infrastructure and workflows. The bottleneck isn’t purely about buying more hardware; it’s that current tools, data practices, and talent bandwidth haven’t kept pace with the scale and complexity of modern materials research.
The most striking finding is how universal the pain points are. Nearly every team reported abandoning projects, expressing concerns about IP protection, and feeling constrained by compute and workflow complexity.
It’s also notable how willing scientists are to accept modest accuracy trade-offs – with the majority (42 %) accepting a 5-10 meV/atom – if it means simulations run up to 100× faster. That signals that the industry is ready for new, more efficient methods.
Fumio Horino, VP Marketing at Matlantis
The Growth of AI in Simulation Workflows
AI-driven simulation is no longer an occasional practice, but a mainstream component of materials R&D.
The report showed that almost half of research teams – approximately 42 % – already use AI-native simulation platforms in production, and an additional third are evaluating or piloting AI-augmented tools.
This shows that AI is being integrated alongside, rather than replacing, traditional methods such as DFT, molecular dynamics, or GUI-based simulation tools.
Respondents estimate that 46 % of their current simulation workload is carried out using AI or machine learning methods.
The remaining half are still relying on conventional physics-based approaches: AI acceptance is significant, but has not yet dominated the R&D sector. Trust and security play a role in this limited adoption.
Trust, Security, and Requirements for AI Adoption
The report shows that 100 % of teams have concerns about IP security when using cloud-based or third-party tools.
Or as Fumio puts it, "the report findings show that 0 % of respondents said intellectual property security is a non-issue".
In other words, every single team has concerns about the security of their data when using external or cloud-based simulation tools. Trust in AI’s accuracy is cautious: only 14 % are “very confident” in the results from AI-accelerated simulations.
Most are moderately confident at best, underscoring a cautious optimism that still needs validation.
Fumio Horino, VP Marketing at Matlantis
Respondents cite a range of necessary assurances before adopting AI-native simulation. These include auditability of training data (28 %), security certifications such as ISO or SOC2 (26 %), regulatory compliance (24 %), legal protections for IP (24 %), clarity on data ownership (23 %), and guarantees against data leakage (22 %).
These preferences suggest that, without strong governance and transparency, even fast or accurate AI tools will encounter barriers to adoption.
For Fumio, the best defense is a combination of architecture and policy. He lists several key steps teams could, or should, implement to safeguard against risk.
Use private cloud or dedicated VPC deployments, enforce strict access controls and encryption, demand full clarity on data ownership and training-data usage, and rely on vendors with strong compliance and audit capabilities.
Zero-trust principles, rotating credentials, validated containers, and clear contractual guarantees are essential. For many organizations, the ideal model is the ability to use the same AI tools in a fully controlled environment – whether on-premises or in a private cloud – so they can gain the benefits without compromising IP security.
Fumio Horino, VP Marketing at Matlantis
Increasing Interest in Specialized Accelerators
Most teams believe that progress in simulation will depend on hardware acceleration beyond standard CPUs, and only about 16 percent think CPUs alone will meet their future computational needs and circumvent the bottlenecks currently hindering simulations.
Roughly 83 % of teams expect to rely on specialized processing hardware, indicating a broad expectation that computational speed improvements will require more advanced architectures.
To try to avoid the speed limitations that R&D teams are facing, Fumio suggests a restructuring approach.
Organizations should redesign workflows around multi-fidelity methods. Fast AI-based surrogates or ML potentials can be used to screen enormous design spaces quickly, reserving expensive DFT or MD calculations only for the most promising candidates. Pre-trained and “out-of-the-box” models reduce the need for constant retraining and help teams reuse work across chemistries.
On the infrastructure side, hybrid approaches – running everyday work on local clusters but bursting to secure cloud resources for large sweeps – dramatically reduce bottlenecks. Workflow automation and caching prevent teams from repeatedly rerunning identical or low-value jobs. And, perhaps most importantly, investing in data organization and standardized metadata ensures that results can be reused, retrained, and shared instead of recomputed.
Combined, these steps let teams explore orders of magnitude more ideas in the same amount of time.
Fumio Horino, VP Marketing at Matlantis
Financial Impact and Simulation ROI
Simulation delivers considerable economic value to organizations. Respondents report an average savings of about $ 109,000 USD per R&D project due to reduced dependence on experimental work.
Approximately one-third state that they save over $ 100,000 USD per project. Most organizations also commit significant annual budgets to simulation. For instance, around half of respondents spend between $ 200,000 and $ 500,000 USD per year, and just under a quarter spend between $ 500,000 and $ 700,000 USD.
The average annual simulation budget is approximately $ 416,000 USD, not an insignificant number.
These figures show that many organizations are investing heavily in simulation as they consistently achieve cost reductions and productivity improvements.
Get all the details: Grab your PDF here!
Future Outlooks
Overall, the report portrays a field moving steadily toward broader AI adoption, but one that is still constrained by speed limitations and security requirements.
Researchers demonstrate a clear interest in tools that can accelerate discovery, as long as they comply with enterprise security expectations.
Simulation already delivers strong financial returns, and teams are prepared to invest further if new platforms can reduce bottlenecks in compute, expertise, and workflow integration.
Reference
Matlantis (2025), Accelerating Discovery: AI Trends in Materials R&D – How Simulation Leaders Are Balancing Speed, Accuracy & Trust. Available at https://matlantis.com/en/resources/useful/accelerating-discovery-ai-trends-in-materials-rd/ (Accessed: 19th November 2025)
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.