Bawendi was an expert in perovskites, but he won the Nobel Prize in 2023 not for his work on perovskites, but for innovations in the chemical synthesis methods of Quantum Dots.
Quantum Dots were tiny semiconductor particles with unique optical and electronic properties, applicable in fields such as LEDs, Infrared Detection, and Solar Cells.
When Bawendi heard that he had chosen perovskites, he smiled. "Light, I'm not sure how skilled you are at experiments, because your paper didn't really showcase your talent in that area.
"Publishing papers in perovskites places extremely high demands on experimental ability. You need to achieve results others cannot, or observe phenomena others cannot observe, and then be able to theorize them."
Bawendi continued, "Of course, if you want to publish in top-tier journals, I know that is important to Chinese students, because you need such credentials to prove your ability and secure better treatment when you return home.
"From that perspective, perovskites really are an excellent direction. Every year, there are no fewer than ten perovskite papers in Nature alone. Add Science to that, and it really is a field where results come easily.
"I'm just not sure how talented you are at experiments."
Chen Yuanguang's confidence soared. Compared to the projects Pangu had given him, this was nothing. "I think I can do it."
Bawendi stared into his eyes for a while, then nodded. "Okay. If you think you can do it, that's good enough. Come on, kid, let's get to work!"
After sending him a list of papers and relevant textbooks, Bawendi told him to study on his own. That was MIT's style—advisors assumed you were a genius.
"Has Chen come to the lab?" Bawendi asked.
An Indian student said, "I haven't seen him."
"All right." Bawendi was somewhat puzzled.
When Chen Yuanguang emailed him again, hoping that he could recommend some new papers, Bawendi replied with an email asking, "Do you have any ideas for the paper yet?"
"I have a rough idea. I plan to study how Machine Learning can be used to predict the high-throughput effects of antisolvents on perovskite stability," Chen Yuanguang replied.
After reading it, Bawendi realized that Chen Yuanguang had still gone down a path familiar to him: computational chemistry.
Machine Learning was good. In 2016, AlphaGo emerged out of nowhere and defeated the ninth-dan Go player Li Shishi, an event hailed as the first year of Machine Learning.
Two years had passed since then, and everyone felt that artificial intelligence needed to be integrated with their own fields, just as internet-plus initiatives had been everywhere twenty years earlier.
Now it was artificial intelligence plus, but AI talent was hard to find. Tech giants like Google, Amazon, and FB offered astronomical salaries to people in the artificial intelligence field, while startups offered one exorbitant price after another.
Finding someone who understood artificial intelligence and chemistry was already incredibly difficult, let alone someone who could combine the two.
Even Bawendi could not find talent in this area. Chen Yuanguang's appearance had given him some ideas, but he had not expected the other man to adapt so quickly. He had found the point of intersection after only a month.
"Light, come to my office tomorrow. Let's discuss the specific research direction," Bawendi said directly over the phone after seeing Chen Yuanguang's reply instead of emailing back.
The next day, Bawendi handed Chen Yuanguang a coffee. "Light, give me a rough outline of your idea."
"My idea is to combine automated characterization, chemical robot synthesis technology, and Machine Learning to explore how the choice of antisolvent affects the intrinsic stability of perovskites.
"For example, we can use different terminal groups to create combinations, such as MAPbI3, CsPbI3, CsPbBr3, and so on, to synthesize several combinatorial libraries. Each library will have its own unique combinations.
"I estimate that we can synthesize more than a thousand compositions in total, then prepare each library twice using two different antisolvents: toluene and chloroform.
"After synthesis, we will automatically perform a Photoluminescence Spectroscopy Analysis every five minutes for a period of time. The appropriate duration will need to be determined experimentally.
"Finally, we will use Non-negative Matrix Factorization to map the time- and composition-dependent photoelectronic properties.
"By applying this workflow to each library, we can identify how the choice of antisolvent affects the intrinsic stability of perovskites.
"In fact, it may be a dynamic process." Chen Yuanguang gave a rough explanation of his idea.
This approach was very similar to the previous directed evolution of proteins, except that it began with experiments, using chemical robots to establish a fixed process before using Machine Learning for analysis.
Bawendi nodded repeatedly as he listened. What a promising prospect. No wonder Levitt had sounded resentful when he called me. "Excellent idea, Light. I support you.
"It will depend on the final results. If they are good, this could be published in Nature or Science.
"If the results are less than ideal, it can still go to JACS or a Nature subsidiary journal. For chemistry scholars who have fallen behind the times, the currently popular artificial intelligence algorithms will still earn some respect."
Chen Yuanguang considered it. The main reason was that internet industry compensation was simply too good. Those capable of combining perovskites and Machine Learning either had not yet matured, since artificial intelligence had only been popular for two years, or they would never pursue a PhD at all—they had long since gone into industry to make big money.
It was only because that kind of money meant nothing to him that he had the opportunity to study here.
"Professor, there's one more thing. Our lab doesn't have a graphics card. I need to buy Nvidia's latest graphics card for the calculations," Chen Yuanguang said.
Bawendi said helplessly, "It seems I've fallen behind too. Buy it. Write me an application form, and after I sign it, take it to Susan and have her procure it."
"Not bad. The results aren't particularly good, but they aren't bad either. Publishing in Nature might be a bit of a stretch, but JACS is more than achievable." Bawendi stared at the results for a long time before speaking.
Chen Yuanguang said, "JACS is enough. While I was working on this project, I also started another one. That project can definitely be published in Nature."
Bawendi was stunned. "Tell me about it."
Chen Yuanguang said, "I built a cross-property deep learning framework on GitHub, mainly for predictive analysis of small materials datasets.
"I mainly used some existing data from the department to train the model. First, I constructed a large dataset, then used that large dataset to build small datasets with different properties, along with their models.
"Through the framework distilled from these models, we can directly input physical properties as its computational and experimental datasets, and ultimately derive the results for its other properties."
Put simply, he had dug one layer deeper beneath the single topic of perovskite prediction and produced a more general result.
Before you continue