Working towards trustworthy AI: advancing high-precision prediction of solar radiation and typhoon intensity

My interest in monozukuri (craftsmanship/making things) dates back to junior high school. When I first got my hands on a computer, I became absorbed in creating my own games and music. This joy of exploration through monozukuri led me to major in engineering at university. My reason for joining Fujitsu Research stemmed from an internship experience. At that time, it was difficult to get information about what corporate research labs actually did, so I participated in the internship with a "just try it out" attitude. The deciding factors for joining Fujitsu were the feeling that I could utilize the programming skills I developed in university to create appealing things, together with being able to envision myself working there.

Since joining the company, I have been involved in natural language processing (NLP) related projects – a technology that analyzes, understands, and generates human language using computers. I have focused particularly on information retrieval and text document analysis, with responsibility for developing technologies to extract key points and summarize information from customer inquiry emails and similar sources. Through the development of these diverse NLP technologies, I have found considerable pleasure in solving everyday problems and improving operational efficiency.

From 2019 to 2021, leveraging the knowledge cultivated in natural language processing, I was responsible for promoting the commercialization of Wide Learning (*1), an Explainable AI (XAI) technology that fuses discovery science and machine learning. This involved explaining the technology at external exhibitions and conducting proof-of-concept experiments with customers. Wide Learning is a highly versatile technology that can contribute to solving problems across a wide range of fields. One of its key features, which has received high praise from many customers, is its ability to explain why a certain result was reached and to identify the factors influencing that outcome. For example, Wide Learning can efficiently uncover the characteristics and combinations of factors that determine election outcomes (winning or losing) from election result data. It can also identify the characteristics of business documents that effectively convey requirements, based on various conditions. In my Wide Learning projects, I focused on devising effective demonstration methods tailored to customers' challenges and inputting appropriate training data into Wide Learning for its application to on-site systems. Steadily achieving results in each project and being able to publish these as academic papers and books has been a great joy for me and a significant encouragement for my future endeavors.

Currently, I am affiliated with the Space Data Frontiers Research Center. This center was established in April 2025, with the aim of contributing to the creation of new space-related businesses. A direct invitation from the head of the center was the catalyst for my transfer. I was approached with the idea that the AI development skills I cultivated during the Wide Learning project could potentially contribute to space weather prediction. Space weather refers to the fluctuating space environment caused by phenomena such as solar radiation, which has various impacts, both large and small, on Earth. Although this is an entirely new field for me, I have found this challenge to be highly appealing—to leverage my expertise as an AI developer and contribute to space development research that protects human lives—and as a result, I decided to transfer. At the current research center, I am responsible for the R&D of solar radiation occurrence prediction using AI(*2). When solar radiation occurs, infrastructure such as artificial satellites, aircraft, and even power plants can be damaged, potentially causing severe impacts like power outages. If the occurrence of radiation can be predicted in advance, effective countermeasures, such as issuing warnings or preparing alternative measures, become possible. For details of our joint research with Tokai National Higher Education and Research System using AI, please refer to the following content (*3).

In addition, as another research activity, I am engaged in R&D for predicting typhoon intensity. This project is being pursued jointly with Yokohama National University (*4) as part of the Fujitsu Small Research Lab (*5) initiative. There are two major challenges in typhoon prediction: track prediction and intensity prediction. The latter is extremely important because if a typhoon rapidly intensifies and approaches Japan, it can cause enormous damage. On the other hand, predicting whether a typhoon will truly strengthen or weaken tomorrow is very difficult. However, since this prediction directly impacts human lives and property, we believe that even a slight increase in prediction accuracy through the use of AI, enabling appropriate preparedness, can significantly reduce damage.

Both predicting solar radiation occurrences and typhoon intensity involve natural phenomena that are difficult to fully understand. For such predictions, not only high accuracy but also the trustworthiness of the results is crucial. Therefore, we are developing AI capable of predicting natural phenomena, based on Wide Learning technology. When dealing with natural phenomena, specialized knowledge is indispensable. For example, to predict the probability of a solar energetic particle event, which is one type of solar radiation phenomenon, it's necessary to extract data indicating connections to the event's occurrence, such as the location, duration, brightness, and historical record of solar flares, and use this as AI training data. For this reason, we are advancing our research in close collaboration with experts from our joint research partners.

However, simply feeding data directly into AI will not yield the desired results. It's essential to input training data that has undergone proper preprocessing and shaping. To obtain reliable results, we ourselves must understand and discern the characteristics of the training data. Repeating hypothesis testing while changing learning parameters one by one, and painstakingly creating appropriate data through trial and error, might seem like tedious work. However, diligently and precisely executing this is extremely important. Our daily research efforts are aimed at "establishing expertise in the application of AI to natural phenomena," and we strive to provide reliable space weather and typhoon intensity predictions that can explain the factors behind their results.

The field of space and natural sciences boasts a diverse array of unique data, including satellite observations and simulations. Handling this data presents challenges akin to grappling with nature itself. Particularly for space weather phenomena, extensive datasets are not always available, requiring intuitive insight to grasp the essence from limited information. We sincerely hope that Takashi, leveraging his data analysis experience cultivated across various fields, will earnestly and flexibly tackle problems that cannot be solved with space knowledge alone, thereby bringing a fresh perspective to space weather forecasting. (Chihiro Mitsuda, Research Director, Space Data Frontiers Research Center)

Titles, numerical values, and proper nouns in this document are those reported when this interview was made.