Tiangong Kaiwu: The Fifth Paradigm

Human civilization's earliest periodization was refreshingly straightforward: the Stone Age, the Bronze Age, the Iron Age. Not divided by political systems, not by schools of thought, but by materials. The discovery of a new material rarely just improved a single product — it lifted entire industrial chains and sometimes defined an entire era.

This is still happening today. In 2011, Ryoji Kanno at the Tokyo Institute of Technology discovered a crystalline material called lithium germanium phosphide sulfide (LGPS). This powdery solid was the first in which lithium ions conducted as fast as they did in liquid electrolytes. Before this, solid-state batteries were an academic backwater; afterward, billions in investment poured in from Toyota, Samsung, and CATL, and solid-state batteries became one of the biggest bets in new energy. Today, this material sells for millions of dollars per ton.

One material, one industry.

But few people know that Kanno spent over twenty years searching within the lithium-phosphorus-sulfur system to find LGPS. His method was the same one material scientists had used for millennia — based on experience and intuition, narrowing the scope, then trying one compound after another.

Today,

AI is trying to change that equation.

The Fifth Paradigm: AI for Science

From Skeptic to Founder

Ziheng Lu, founder of Kairosmaterials, can pinpoint exactly when his conviction in this began.

He studied mechanics at Nanjing University of Science and Technology, then pursued his PhD at HKUST doing computational modeling of solid-state electrolytes. When applying, he wrote a research proposal that now seems almost naive: using molecular simulation to model everything in the universe. His advisor Francesco read it, smiled, and said it sounded nice.

"Only later did I realize the scale was impossibly large," Lu said with a wry laugh.

This experience seeded persistent skepticism about computational methods. After two years of simulation work, he felt it was "too detached — I'd never even seen what real materials looked like," so he went to Yale to learn the full experimental toolkit, returned to Hong Kong to build a lab, and spent the next several years primarily doing experiments. After graduating, he spent three years as a faculty member at the Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, supervising ten master's students while trying to scale his lab's results toward commercial production.

During that phase, he was "fairly disappointed" with computation and AI.

Material science has an uncomfortable reality: for decades, the field's core advances often depended on the intuition of a handful of genius scientists. Of the most important cathode materials for lithium batteries, at least two were directly proposed by Goodenough (LiCoO₂, LiFePO₄), while other systems (such as LiMn₂O₄, lithium-rich manganese-based materials) were built on the theoretical framework of transition metal oxides he established.

"I don't know why he was so smart," Lu said. "But that's exactly what we're trying to change. I'm not that smart, but I want AI to be."

Democratizing genius — it's a proposition that sounds grandiose but is actually quite concrete.

The turning point came at Cambridge. From 2020 to 2022, he did postdoctoral work there, using an algorithm called random structure search combined with small-scale AI to find inexpensive lithium-ion cathode materials. They actually found one — a manganese-containing oxalate cathode material. Its performance wasn't spectacular, but it was a completely new compound, and it was actually synthesized in the lab.

Ab initio random structure searching for battery cathode materials

"That was the first time I felt that computational work could actually deliver. And that AI was useful."

In 2022 he joined Microsoft Research Asia with a clear goal: scale up the model. In the period before GPT-3.5's release, AI for science was a heavily anticipated direction — computer vision had been exhausted, language models hadn't yet taken off, and both academia and industry needed a new narrative. Lu led a team to train a model called MatterSim — given the positions and types of atoms, predict the energy and forces between them, and the various physical properties derived from these.

MatterSim prediction comparison chart

What he wanted to do traced directly back to that naive idea from his PhD application: one model that could calculate any material.

Nobody on the team believed it would work.

At the time, a senior condensed matter physicist gave Lu a simple challenge: never mind everything in the universe — just pick a few materials and calculate their phonon spectra accurately. This is an undergraduate-level concept in basic physics, corresponding macroscopically to heat capacity.

"If you can't even do this, don't talk about anything else."

"Honestly, I had my own doubts," Lu admitted. "But I told everyone: 100% it will work, absolutely."

They started feeding data into the model and expanding its parameters — essentially the same path GPT had validated: scaling.

Until one day, they decided to test it.

"After the test, nobody argued anymore. Nobody said anything. Everyone just kept their heads down and kept working."

The model surpassed all small models specifically designed for this property on a standard benchmark called MatBench. This meant a general, universal model for all materials was actually outperforming tailored, specialized models on specific tasks. The same logic playing out in language models — sufficiently large general models eventually defeat all specialized ones.

This became one of the foundational sources of confidence for Lu's decision to start a company. In 2025, he left Microsoft and founded Kairosmaterials.

The company's name is taken from Song Yingxing's Tiangong Kaiwu (The Exploitation of the Works of Nature) from the Ming Dynasty — China's first systematic encyclopedia recording materials and craftsmanship, covering virtually all manufacturing knowledge from iron smelting to ceramics to textiles.

Why It's Different

There's a large gap between confidence and reality.

AI for materials isn't a new concept. Many companies have been working in this space for years, with significant academic impact but slow commercial progress. These companies, domestic and international, have taken different approaches — some leaning toward pure algorithms, some toward automated labs, some betting on quantum chemical calculations. But they all address the same fundamental question: can general-purpose large models actually work for materials discovery?

Google DeepMind's GNoME predicted 2.4 million potentially stable new materials in 2023, but subsequent synthesis validation found many predictions unreliable. A recurring plotline in this field: AI predicts an amazing material, then the lab can't make it. The computational people say "your experimental capabilities are lacking," the experimental people say "your predictions aren't reliable."

What's the root of this conflict?

Material science has a fundamentally different challenge from drug discovery: solid-state synthesis. In pharmaceutical labs, most operations involve liquid handling — mixing, dripping, stirring — things that high-throughput robots already do well. But materials require mixing inorganic powders in precise ratios, then sintering at specific temperatures and pressures — processes far more complex than liquid operations, and much less automated.

Typical solid-state synthesis workflow

In other words, even if AI's predictions are accurate, there's still a long road from prediction to synthesis to validation to mass production — one where AI currently can't help much.

Lu knows this intimately. He straddles both computational and experimental worlds — learned synthesis techniques at Yale, did industrial scale-up in Shenzhen, and personally synthesized AI-predicted new materials at Cambridge. So from day one, his company built both AI and experimental teams simultaneously, and insisted they couldn't be separated — "It has to be one team. AI people need at least intermediate-level condensed matter physics knowledge. Senior experimentalists need systematic understanding of AI. Complete linguistic interoperability."

The磨合 mechanism was equally unpretentious: two teams, required to eat lunch together every day. "Just chatting over lunch, that solves it."

This integration produced concrete effects: the model predicts candidate materials, the neighboring lab starts synthesis validation the next day. When it works, data feeds back to make the model more accurate; when it doesn't, both sides sit down to diagnose the cause, adjust, and try again. The feedback flywheel spins faster.

Additionally, AI for materials differs fundamentally from large language models. Language model training data is fixed — the entire internet's text corpus is just there, and what you need is to train a sufficiently large model to digest it. But materials has no such ready-made database. The materials data accumulated in human history is negligible compared to internet corpora.

"So for us, the bottleneck is more in data, not models."

This means Kairosmaterials must do more than train models — it must create data. They use quantum chemical calculations (DFT) to generate large amounts of synthetic data, while producing experimental data from their own lab. Model training and data production proceed in parallel — the model predicts candidate materials, the lab validates, and validation results become new training data.

"It's the same logic as embodied intelligence," he said. "Embodied intelligence didn't have that much real-world data at first, so they used simulation first. Later they found a sim-to-real gap, and started generating data in the real world. Materials are the same — you need to first create an environment where data can scale, then scale the model alongside it."

The sim-to-real pathway in embodied intelligence

From Twenty Years to Six Months

"If we put Kairosmaterials' model back in 2011, could you have found LGPS faster than Kanno?" We posed this hypothetical to Lu.

"I think most likely yes, and significantly faster." He paused. "Won't claim absolute certainty, but I'd estimate around six months. He spent over twenty years."

During his Cambridge period, he had done similar retrospective tests — feeding the elemental combinations of several known important solid-state electrolytes into the model to see if it could "rediscover" them. The results were positive. Lithium lanthanum zirconium oxide — a solid-state electrolyte second only to LGPS, with an extremely complex structure — could also be found.

Schematic of LLZO crystal structure rediscoverable in retrospective testing

"This is why we believe AI is genuinely useful — things that couldn't be found with traditional methods before, AI-assisted methods can now find."

Of course, retrospective validation and genuine frontier discovery are different things. Finding known good materials proves the model has capability. What really excites the field is: can it find good materials that humans don't yet know about?

Lu's answer: "AI has already found some useful new materials, including ones we've found ourselves. But that truly heavyweight, industry-defining discovery? Definitely not yet."

He sees this as a matter of time, not direction. "The technical preconditions are already in place. Scaling works for materials prediction. General models are starting to surpass specialized ones. Our lab can rapidly validate predictions. What's needed now is time — to search through sufficiently large chemical space, to run through the full pipeline from discovery to validation to mass production."

Kairosmaterials' next directions fall into two categories.

One is "make serious money": energy is the biggest theme. AI's underlying driver is itself energy, and ultimately competes on energy — energy storage, power batteries, and energy conversion materials all fall in this scope. Further out on the periphery are "small materials, big leverage" situations stuck in supply chains — for example, the inner walls of spherical LNG cargo tank compartments on oceangoing vessels, which require a special adhesive stable at extreme low temperatures and pressures, with extremely high unit prices, long dominated by a few foreign companies. Categories with clear technical bottlenecks and supply chain vulnerability are where AI's search capabilities can most easily切入.

The other is "show muscle": superconductors are the classic example, whether room-temperature or even ambient-pressure superconducting materials would be the holy grail that condensed matter physics has pursued for decades; first-wall materials for nuclear fusion — alloys or ceramics that can withstand hundred-million-degree plasma impacts inside fusion reactors. These may not make money in the near term, but if achieved, would redefine humanity's energy boundaries.

On business model, Kairosmaterials is clear: first run through a complete closed loop themselves — AI discovers materials, lab synthesizes and validates, scale-up production, deliver to customers. Lu draws an analogy to Flagship Pioneering — the biotech fund that incubated compounds from early stage to standalone companies, producing firms like Moderna. He hopes Kairosmaterials can forge a similar path in materials.

This isn't cheap. Today's two biggest cost centers are compute and AI talent, especially the latter — significantly more expensive than comparable talent in other fields.

Monolith led Kairosmaterials' angel+ round. Lu said his first impression walking into Monolith's office was striking: "It felt completely different from other investment institutions. Very minimalist, all glass transparent, very young people inside — it didn't feel like an investment firm, more like a tech startup." Additionally, communication with Monolith founder Xi Cao was surprisingly smooth, with quick alignment on technical direction and business judgment. The entire decision process was clean and direct. "The most efficient communication I've experienced with any institution."

A Paradigm-Level Leap

Looking back at history, humanity's way of understanding the material world has undergone only four fundamental transformations:

First was empirical trial and error — how materials were discovered for millennia;

Second was theory-driven — thermodynamics and quantum mechanics let humans begin describing material laws mathematically;

Third was computational simulation — with computers, scientists could do some experiments on screens;

Fourth was data-driven — when data volume reached sufficient scale, statistics itself could reveal patterns.

Material science is an extremely ancient yet extremely important field. Humanity's hunger for new materials has never ceased — higher energy density batteries, more efficient solar cells, more heat-resistant alloys, cheaper catalysts. If found, each of these could撬动 a hundred-billion-dollar market.

Now, AI offers a fundamentally different possibility: no longer relying on the intuition of a few geniuses, but using models to search systematically through all of chemical space. What it changes isn't just efficiency, but the paradigm itself. The fifth paradigm is arriving. This is what companies like Kairosmaterials represent.

The Fifth Paradigm: AI for Science

The deeper significance of AI for science — not just making scientists work faster, but making previously impossible things possible. Material science, drug discovery, protein structure prediction — all follow this pattern. When AI begins intervening in domains of deepest human intellect, largest search spaces, and most fundamental civilizational impact, what it brings is no longer incremental improvement, but paradigm-level leaps.

What will be the next era-defining material? Perhaps an electrolyte that finally makes solid-state batteries mass-producible, perhaps a photovoltaic material that doubles solar efficiency, perhaps some entirely new substance we can't yet imagine — hiding in some corner of chemical space that human intuition could never reach.

Previously, finding it required a genius scientist spending twenty years.

Soon, perhaps it won't.

Kairosmaterials was founded in 2025 by former Microsoft Research Asia researcher Ziheng Lu. The company name derives from the Ming Dynasty scientific classic Tiangong Kaiwu (The Exploitation of the Works of Nature). It has completed an angel+ round led by Monolith. The team is actively recruiting: AI researchers with machine learning and graph network/generative model backgrounds, computational materials scientists with quantum chemical calculation or condensed matter physics experience, and senior synthesis experimental experts with industrial scale-up experience.

If you believe AI will transform materials science, welcome to contact careers@kairosmaterials.com

Note