AI startups tackle physical challenges in materials discovery

In the tech world, a new wave of startups is putting artificial intelligence to work on physical challenges, aiming to invent materials that perform better, cost less and meet critical industrial demands. These firms are training AI models on scientific literature and experimental results to simulate chemical and physical properties, while some have built self-driving labs that can autonomously create and test new formulations. Their early targets include high-tech sectors like data centers, quantum computing and nuclear fusion, where advanced materials are essential but often difficult to produce.

AI-Driven Materials Discovery

Periodic Labs, launched in September 2025, is among the pioneers in this space. The startup aims to generate completely new materials, though scaling production can take years. Periodic has already identified potential applications in magnets, semiconductors and superconductors, all critical for high-tech industries.

The firm’s Montreal team includes researchers from Google DeepMind, Meta and ServiceNow, bringing expertise in AI breakthroughs. Fedus, a former OpenAI employee who helped develop ChatGPT, earned his doctorate under AI pioneers Yoshua Bengio and Hugo Larochelle at Université de Montréal. He emphasizes that academic literature alone isn’t sufficient for training AI in chemistry and physics, as papers often omit failed experiments and present discoveries in a simplified manner. To address this, Periodic uses autonomous labs to gather real-world data, creating a feedback loop between digital simulations and physical testing. The company charges clients for software access and may share intellectual property revenue from joint R&D projects.

In October 2025, Periodic raised a US$300 million seed round from investors including Radical Ventures, Accel, Andreessen Horowitz and Khosla Ventures. The technology addresses a key challenge: ensuring AI-suggested materials can actually be manufactured using common processes or re-engineered methods. Fedus notes that clients lack time to redesign their operations, so Periodic works within existing infrastructure while solving physical and engineering barriers to production.

Racing to Cool Data Centers

Across Toronto, FL01 is focusing on thermal fluids for data center cooling, a sector where manual testing is slow and labor-intensive. CEO Mohammad Zargartalebi, who co-founded the company in November 2025 after developing sensors during a University of Toronto fellowship, partnered with BP to build a robotic scientist that formulates and tests coolants in minutes. “The properties of coolants are not easily predicted,” he said, explaining that starting with hardware gave FL01 the data needed to train its AI model. Working with BP, the team has identified 30 potential formulations, three of which passed stability, corrosion and other tests and are ready for commercialization.

FL01 sells testing services and may eventually license or manufacture its inventions. Despite being smaller and less funded than competitors like Periodic, Zargartalebi argues its narrow focus offers an edge.

In the same accelerator cohort, Meissner is developing AI-designed superconductors for quantum computing and nuclear fusion. CEO Olivia Leng explains that general materials platforms struggle with exotic materials like superconductors. Meissner focuses on devices made from readily available ingredients using existing processes. Last month, the company announced a $3.6 million pre-seed round from BDC Capital and deep-tech executives, aiming to create superconductors that operate without extreme cooling.

Closing the Theory-Practice Gap

Periodic Labs addresses a fundamental challenge in scientific research: academic papers often omit failed experiments and present discoveries in a simplified manner. CEO Liam Fedus explains that this creates a “fictionalized view of the scientific method” that hampers AI training. To overcome this limitation, the company employs autonomous labs that conduct real-world experiments, generating the missing data needed to train accurate models. This approach creates a feedback loop where digital simulations guide physical testing, and experimental results refine the AI’s predictions. The system effectively allows the technology to “do the scientific method” and gain deeper insight into materials engineering.

Focus and Funding Challenges

While FL01 operates with limited resources—having graduated from the Creative Destruction Lab accelerator without raising venture capital—it leverages its narrow specialization as a competitive advantage. CEO Mohammad Zargartalebi argues that this focused approach provides an edge over broader platforms like Periodic, which are much larger and better-funded but spread their efforts across multiple material categories.