AI is Powering the Future of Material Science: From Lab to Real-World Breakthroughs

13 min readOct 29, 2024

Is the future of material discovery already here? Imagine if we could design novel materials for carbon capture or battery storage before even stepping into a lab — faster, cheaper, and with pinpoint accuracy. This is the promise of AI in material science. In this blog post, we explore how advancements like Physics AI and generative models are unlocking new opportunities, from revolutionizing energy storage to addressing climate change. Discover how startups, corporates, and investors are working together to bridge the gap between theoretical innovation and real-world application, and the challenges they must overcome to achieve commercial success.

Material science has fueled technological leaps, from semiconductors to renewable energy. However, progress has often been slow, constrained by the laborious and resource-intensive process of trial-and-error experimentation. Today, however, the fusion of artificial intelligence (AI) and material science promises to change that, enabling faster, more accurate discoveries and reducing the gap between theoretical innovation and real-world application.

At Hitachi Ventures and KOMPAS VC we have been at the forefront of scouting startups in this exciting intersection. Our research has given us deep insights into the transformative potential of AI, particularly Physics AI and materials discovery. This blog post shares our key takeaways, from the promise of AI to the challenges in scaling and commercialization. We encourage you (founder, investors and corporates) to join us in an ongoing dialogue as we deepen our collective understanding.

The Evolution of Material Discovery: A Slow, Expensive Process

Historically, material discovery has been a painstaking and resource-intensive process, relying heavily on laboratory experimentation. Each new material typically took years, if not decades, to discover, validate, and bring to market. The cost was immense, and the success rate low. The traditional model, though effective in certain breakthroughs like semiconductors or high-strength polymers, lacked scalability and speed — two critical elements for addressing the challenges of today’s industries, such as energy storage, carbon capture, and sustainable manufacturing.

The rise of AI-driven material discovery represents a dramatic shift. Rather than relying solely on experimentation, researchers can now use machine learning models to predict the properties of materials before they are ever synthesized in a lab, scaling their innovative capabilities. This approach, known as materials informatics, uses AI algorithms such as Graph Neural Networks (GNNs) and Physics-Informed Neural Networks (PINNs) to analyze massive datasets, identify patterns, and generate predictions about material behavior at the atomic level.

Why do the recent advancements in AI-driven material discovery matter?

AI innovations in material sciences are crucial because they accelerate discovery, enhance sustainability, optimize manufacturing, and address global challenges such as energy efficiency, climate change and medical advancements. By combining AI’s data-driven insights with material science’s experimental techniques, researchers can create more effective, sustainable, and customized materials that will shape the future of industries worldwide and address broader societal needs. For example, it can be argued that the technologies needed to achieve Net Zero by 2050 have yet to be developed and achieving climate neutrality hinges on our capacity to innovate. Unfortunately, the current pace of innovation falls short of what is necessary to meet these net-zero goals. Fields like chemistry or particle physics have such a vast combinatory space that without AI support, humans will never be able to fully explore them and unlock synergies that were previously inaccessible.

**Almost any natural science and engineering field will benefit from AI advancements in material discovery**

Physics AI vs. Physical AI: Understanding the Distinctions

Through our research, we’ve recognized a key distinction between Physics AI and Physical AI. Both approaches are instrumental in driving innovation within material science, but they tackle different aspects of the discovery process.

When discussing foundation models in scientific material science discovery, it’s important to highlight that large language models (LLMs), in their current form, are not directly applicable to scientific discovery. However, certain properties of LLMs are highly beneficial for this field:

Natural language generation and comprehension: LLMs excel at understanding and generating human language, which facilitates seamless interaction and allows scientists to prompt and guide the models easily.
Reasoning abilities: While reasoning is a crucial aspect of scientific discovery, LLMs are still developing this capability, which is key to progress in the field.

However, scientific discovery, especially in the natural sciences, requires precise calculations — an area where LLMs still fall short. Additionally, all discoveries must be verified through experimental testing, adding another layer of complexity.

A significant challenge in applying LLMs to material science is the scarcity of scientific data, in contrast to the abundant web data used to train many LLMs. Scientific data can be generated synthetically or through experiments, but we also have centuries of accumulated knowledge, expressed in precise mathematical forms, such as differential equations (e.g., quantum physics), that describe the laws of physics with great accuracy.

In material science, there is a critical tradeoff between the availability of data and the need for inductive bias. Simply feeding models historical experimental data is insufficient. When data is scarce, as is often the case in scientific discovery, compensating with strong inductive biases becomes crucial. These biases are informed by the specific context and the intended application, enabling models to operate effectively despite limited data. Striking this balance is key to advancing our understanding of complex scientific phenomena.

Physics AI is primarily focused on understanding and simulating the fundamental laws of physics. With advanced AI models, researchers can simulate material behavior under different conditions, accelerating the identification of promising candidates for testing. Physics-Informed Neural Networks (PINNs), for example, integrate physical laws into their models to provide high-fidelity simulations that predict material properties. These models are particularly useful in fields like nuclear fusion, where accurate simulations can reduce the need for costly and time-consuming experiments.
Physical AI, in contrast, involves systems that interact with the physical world. In material science, it could involve robotic solutions, smart sensors and automated laboratories that run real-time experiments and adjust parameters autonomously, further accelerating the discovery process. A key challenge is to find the optimal miniaturization for a given problem space while keeping the real-world application in mind.

Integrating both, Physics and Physical AI, in an end-to-end closed loop that on one side simulates materials’ behavior to identify the best candidates and on the other runs early-stage experimentation from which all metadata is gathered and fed back to the Physics AI, could lead to a powerful active learning process. However, this is a herculean task that would require significant knowledge and expertise of multiple domains including data science, AI, physics, chemistry, robotics and engineering.

Emerging Techniques: Diffusion Models and Generative AI

Among the most exciting developments in AI for material science are diffusion models and generative AI models. Both approaches have rapidly gained traction among startups and incumbents aiming to accelerate material discovery.

Generative AI models and GNNs (graph neural networks), such as those used by DeepMind’s GNoME, go beyond prediction. These models suggest screen and search for novel compositions with embedded DFT (Density Functional Theory) simulations in the loop. GNoME leverages GNNs to model materials at the atomic level. Some of these materials have never been synthesized before. GNoME, for example, has already identified 380,000 new stable materials, including potential superconductors, which could revolutionize industries ranging from quantum computing to energy storage. Doing so GNoME can explore a much larger design space than traditional methods, vastly increasing the potential for novel discoveries.
Diffusion models, a subcategory of Gen AI model, are computational frameworks used to simulate the spread of particles (heat transfer), energy, or other quantities through a medium (mass transport, and the spread of pollutants). These have proven especially valuable for generating synthetic data, which is crucial for training AI models in environments where large-scale, high-quality datasets are scarce. For instance, in fields like carbon capture or battery materials, diffusion models help generate new chemical compositions and simulate their behaviors before conducting real-world experiments. A notable example is Microsoft’s MatterGen, which exemplifies this approach.
PINNs (not the most recent technique as initial paper published in 2019) incorporate physical laws into the neural network training process to confirm the behavior of physical systems governed by physical laws (i.e. temperature distribution through the materials). PINNs measure a step loss function, assessing and ensuring the model adheres to the underlying physics (which is used in several advanced simulations these days).

The Challenges: Data and IP strategy, Computational resources, Innovation pace, and Scalability

While the potential for AI in material science is immense, significant challenges remain. Our working hypothesis is that the slow pace of innovation in material science limits AI-driven breakthroughs from achieving broad commercial success.[pa8] The industry tends to be incremental in its innovations, with major breakthroughs occurring only once in a decade or more. This slow cycle is further complicated by the high cost and complexity of scaling new materials from laboratory experiments to full-scale production.

Data bottlenecks are one of the most significant hurdles. Access to unique datasets is essential for training AI models to predict material properties accurately. However, these datasets are often proprietary and hard to come by. Without access to exclusive, high-quality data, it becomes difficult for startups to differentiate themselves from incumbents or each other. This is where corporates like Hitachi and others can play a key role. By partnering with startups and sharing proprietary data, corporations can help startups train their AI models while retaining control over the intellectual property. The IP strategy in itself is also a challenge depending on the source of data, the discovery, the material synthesis process, the manufacturing control and end customer demands. Startups need to spend significant time and expenses on laying out a comprehensive IP strategy before entering into any agreements with data providers or end customers (which are at times one and the same).
The computational resources and expertise needed to create detailed models of real-world phenomena are significant and often come at a high cost. Material science datasets are often vast and diverse, containing millions of molecular structures, quantum properties, thermodynamic behaviors and environmental factors, which demand significant storage and computational resources. In addition, optimizing materials involves solving problems in a high-dimensional space through advanced simulation techniques (quantum simulations, DFT etc.) with high precision. These demands necessitate powerful hardware like GPUs, supercomputers, and high-performance computing clusters to efficiently carry out the calculations and optimizations needed.

**GPU costs have dropped, leading to cheaper model training**

The 75% drop in GPU costs over the past year makes it more affordable for startups to scale model training and complex computations. This is especially beneficial for fields like material science, where vast datasets and advanced simulations require significant computational power, and it’s securely a very positive tailwind for accelerating innovation.

Scaling up technology is another critical challenge. Once a material is discovered in the lab, scaling it to commercial production can be fraught with difficulty. The process of manufacturing materials at scale often requires new production techniques, which can be costly and time-consuming to develop. This is especially true in industries like consumer electronics and automotive manufacturing, where integrating a new material into existing supply chains can take years. Similarly, application and formulations considerations are crucial to ensure material discoveries translate to real-world cases and can be applied in existing or novel methods of use. Corporates can and need to be an active supporter, reducing the gap between discoveries and scaled utilization by bringing in their years of experience in scaling and applying solutions for customers.
The cost of innovation remains high. The materials industry is notoriously cost-sensitive, with new materials needing not only to perform better but also to be cost-competitive. AI may help reduce the time to market, but it cannot always overcome the dual challenge of performance and cost-effectiveness. For any company in this space choosing the right market is critical, where some markets are non-starters by themselves, or need significant technological innovation to being able compete against incumbents on a feasible investment timeframe. Discovering a novel ingredient which enables new use-cases and businesses might be riskier, but competing against an existing market and efficient industry like the chemical industry is something which needs to be assessed sufficiently, to ensure target prices at realistic volumes can be met.

Startup Landscape: Companies to Watch

While the space remains challenging, we’ve engaged in discussions with many exciting startups that are pushing the boundaries of material discovery through AI. These companies are exploring various use cases, from designing new materials for carbon capture to revolutionizing battery technologies.

Below is a snapshot of the startups we’ve been following closely:

These companies represent just a few examples of the growing number of startups entering the material science space. Many of them are employing unique strategies to differentiate themselves from incumbents, particularly by focusing on data-driven approaches and scalable AI models.

While the materials informatics space is constantly evolving, Citrine and Noble AI stand out for their advanced science-based AI capabilities, surpassing traditional materials platforms. With more companies focused on building long term moats, Cusp AIand Orbital Materials are two companies building their own proprietary datasets and foundation models with an initial focus on materials for carbon capture. Lab automation is advancing through purely cloud based software solutions from companies like Alchemy (digital lab) and Uncountable (lab management) while Dunia and Chemify are taking this further by integrating hardware and software for comprehensive R&D workflow automation and accelerating the vision of fully autonomous labs. We are also seeing end-application specific formulation development companies like Polymerize (for polymers) and Phaseshift (for alloys). Entalpic and AutomatSoln (specializing in battery formulations) offer validation and testing services while more vertically integrated companies like Chemix (novel batteries) and C1 — Circular Carbon (green methanol) are managing the entire process from formulation to large-scale manufacturing.

In terms of enabling technology, companies such as MQS and Quanscient are applying quantum simulations to significantly reduce time-to-market and computational resources needed in material development. We expect to see more companies emerge that will use novel compute architectures and simulation methodologies to both democratize and radically improve the efficiency of the processes.

Strategic Options for Corporates: Collaboration and Open Source

For corporates like Hitachi, the question becomes: How can we actively participate in this AI-driven material science revolution? Based on our findings, we see several promising options:

Collaborative Research: Corporates can engage in open-source research consortia, contributing proprietary data to train AI models. This not only accelerates material discovery but also allows corporates to gain insights from a broader range of research initiatives. For example, Meta has partnered with several startups in the material science space, providing data on carbon capture to train AI models. Explore the Open Catalyst Project. Another good example is IBM RXN in Zurich where several companies from the chemical/pharma space jointly explore and advance the area of lab automation to enable faster proprietary data generation, which is a key bottleneck in chemistry, to train the models from the previous chapter. More details on IBM RXN for Chemistry.
Becoming a Design Partner: Corporates can partner with startups to provide proprietary data for training AI models while retaining ownership of the resulting IP. This approach allows corporations to drive innovation in material science without bearing the full cost and risk of in-house R&D. Building a trusting relationship is key, especially with new partners, to ensure the IP related to the data is kept safe.
Computational resources: Leading companies providing high-performance computing (HPC), cloud platforms, and specialized AI tools to accelerate research and development include Microsoft (Azure Quantum and Azure HPC), IBM (Quantum), NVIDIA (Clara Discovery and Omniverse) and Schrödinger (Material Science Suit) among others. Some of them have started collaborations with research labs to provide computing power, however they could also explore striking partnerships with startups, providing computational resources at lower costs.
Manufacturing resources: Partners in the value chain can offer scale-up facilities and their experience scaling products, making it possible for novel discoveries to reach relevant volumes. One example could be a three-way-partnership between a startup developing novel technologies, an application focused corporate with their customer base and a volume scale up partner with existing production capacity, in which all three parties benefit from the knowledge exchange and advancement of their field.

Conclusion: A Path Forward for AI and Material Science

The intersection of AI and material science holds the potential to unlock unprecedented innovations. However, the path forward isn’t without challenges. Data bottlenecks, scaling issues, and the high cost of innovation remain major hurdles. Yet, through strategic collaborations and leveraging proprietary data, corporates like Hitachi and VC platforms like Kompas can play a pivotal role in overcoming these obstacles, by facilitating connections between the right startups and corporates.

At Hitachi Ventures (Galina Sagan and Vamsi Patti) and KOMPAS VC (Ilena Mece), we remain excited about the possibilities and look forward to close collaboration with both sides driving the materials discovery development further. The combination of AI innovation, corporate partnerships, and unique data access will be the key drivers in shaping the future of material discovery. Though we have yet to invest, we continue to monitor the space closely and believe that the next wave of breakthroughs is just around the corner.

As AI continues to push the boundaries of material science, what role do you think collaboration between startups, corporates, and investors will play in overcoming the challenges of scaling and commercialization?

About Hitachi Ventures
Hitachi Ventures is the strategic global venture capital arm of Hitachi Group, a global industrial player active in a broad range of technology sectors. Hitachi Ventures invests in innovative companies that address society’s key technological, social and environmental challenges in areas like Mobility, Smart Life & Health Industry, Energy & Environment, IT, New Frontiers. Hitachi intends to leverage its strong position in multiple global technology markets, its expertise and network to support and promote Hitachi Ventures portfolio companies.
For more information, please visit www.hitachi-ventures.com/

About KOMPAS VC
KOMPAS is a specialist early-stage venture capital firm with offices in Amsterdam, Berlin, Copenhagen, and Tel Aviv. We back innovative and scalable technology solutions that reduce emissions, increase productivity, and strengthen business resilience in the built environment and the manufacturing industry. Fund I (€135m) targets late Seed and Series B hardware and technology companies in Europe, UK, Israel and the US.
For more information, please visit www.kompas.vc