黃建中's 火種源 platform aims to reconnect fabrication speed with data and the next decision

Thirty-six cells in five hours. For a laboratory where graduate students still fabricate devices one by one and organize each dataset by hand, that is already an impressive pace. 黃建中 immediately followed it with another number: even a heavily simplified materials problem can leave roughly 1.12 million combinations from only a handful of composition ratios.

Put the numbers together and the limit of high throughput appears. Equipment can shorten the time per device, but it cannot fabricate all 1.12 million possibilities. 火種源 is meant to address what happens after the 36 cells are complete: which results can be trusted, and where the next set of conditions should come from.

黃建中 said conventional high-throughput work still contains a broken junction between one completed experiment and the next. The gap appears before any model begins to calculate. Formulations, equipment settings and measurements are copied by hand. Good samples enter reports; failed ones are easily set aside. A team preserves a few successful points while the surrounding dead ends disappear. When the person or batch changes, the next researcher often pays the same tuition again.

火種源 links fabrication, transfer, measurement and sample history. After one round, a model proposes candidate conditions for the next. Those candidates must first know what product they are intended to serve. A rooftop power plant prioritizes efficiency and lifetime. Building integration also has to handle transparency, color and area uniformity. Indoor low-light and agricultural applications have different spectral needs. Optimizing only the highest conversion efficiency may produce an answer that cannot enter a product.

Whether the loop can learn first depends on whether data from two rounds can be compared. The 火種源 system presented by 黃建中 encloses processing in a sealed space or glovebox; the water and oxygen condition stated in the talk was 0.01 ppm. When substrates move between tools and need to be turned over, they pass through enclosed transfer and flipping chambers without first leaving the controlled or vacuum environment. Otherwise, the difference between two devices may be nothing more than the environment one of them encountered during transfer.

黃建中 noted that a year earlier the forum had seen only the equipment design, whereas this talk included photographs of the physical system. Moving from a drawing to hardware gave the enclosed transfer and flipping chambers a concrete form. Whether the platform can operate steadily over time still has to be answered by continuous data.

The platform also keeps a minute-by-minute log for every tool. When a sample deviates, the team can trace what happened at each process step. A history does not guarantee that every anomaly will reveal its cause. It does, however, turn the sample record from a post-experiment archive into a clue for investigating material, equipment and process events. “Thirty-six cells in five hours” gains another meaning: the batch leaves experience that can enter the next round, not merely 36 results.

Scaling area will test that experience again. The current 火種源 high-throughput process mainly uses smaller glass substrates and spin coating. 黃建中 identified A4, roughly 30-centimeter-square substrates, modules and encapsulation as later validation points. As area grows, uniformity, yield and reliability enter the score. If a champion formulation on a small cell carries neither its process history nor the boundary of failed conditions, the next size may still require starting over.

There are already research examples in which a machine selects the conditions for a later round. A 2026 Nature paper connected machine-learning-guided molecular design, Bayesian optimization and automated device fabrication in a closed loop, extending from small cells to mini-modules. It provides an external reference: results from one round of perovskite experiments can inform the next decision. The forum presentation of 火種源 focused on platform architecture, controlled transfer and later scale-up points. What it now has to accumulate is continuous operation, cross-batch results and the judgment that survives the move from small cells to larger formats.

黃建中 condensed the shared-validation idea into a shift from “buying equipment” to “buying R&D efficiency.” Materials suppliers, startups and universities do not all need to own an automated line. If the same material can enter a shared controlled process and carry its failed experiments into scale-up validation, equipment time can become time shared across an R&D network. Whether that service can work still depends on real collaborations and cross-batch data. 火種源's contribution in this talk was to move the finish line of high throughput a preliminary half-step farther—from completing a batch to letting the next begin without returning to zero.

Sources and further reading

Talk and organizer material

Primary paper