Interviews with Faculty Researchers

– Interview with Prof. Yang Ming
Assistant Professor, Department of Applied Physics

The discovery of new functional materials has traditionally relied on time-consuming and costly trial-and-error methods, often taking over 20 years for a material to move from initial discovery to commercial use. Prof. Yang Ming is transforming this process through a data-driven, AI-powered approach that significantly increases the speed, accuracy, and efficiency of identifying advanced materials for electronics and energy technologies.

Unlike traditional databases or search engines that passively retrieve past results, AI-driven models actively learn from large datasets, simulate material behaviours, generate hypotheses and optimize experimental parameters. This allows researchers not only to explore existing knowledge but also to predict new materials and uncover hidden patterns.

Prof. Yang’s research leverages high-throughput first-principles calculations—automated, quantum mechanics-based simulations that evaluate materials without needing physical experiments. In a project focused on high-k dielectric materials for next-generation 2D electronics, his team began with over 140,000 known compounds. By filtering these using key factors like band gap and dielectric constant, they identified about 1,000 promising candidates. Further semi-automated large-scale simulations narrowed the list to around 20 high-performance dielectric materials for 2D semiconductors. This process is estimated to be 4 times faster than conventional methods.

A major innovation in Prof. Yang’s work is the use of physics-informed machine learning, where physical laws are embedded directly into AI models. This enhances accuracy, reduces reliance on large datasets, lowers energy consumption and improves model transparency. His team recently encodes short-range interaction into AI model, in which only local structures are used for the graph representation, making them especially effective for predicting complex material properties such as adsorption and defect behaviour.

Despite the breakthroughs, challenges remain, particularly the need for greater computing power and smarter algorithms to handle vast material datasets. However, with advances in GPUs, parallel computing, and techniques like surrogate modelling and active learning, the pace of discovery continues to accelerate.

By integrating AI, physics and vast material databases, Prof. Yang is reshaping how new materials are discovered. His research supports faster innovation, reduced costs and sustainable development, while positioning Hong Kong as a leading centre for AI-driven materials science.

加速功能材料創新：以人工智能與數據推動先進電子科技發展

– 楊明教授專訪
應用物理學系助理教授

傳統功能材料的研發往往依靠耗時且成本高昂的反覆試驗法，令研發由初步發現到實際應用可能需時超過20年。為此，楊明教授利用大數據及人工智能，為電子與能源技術研發尋找先進材料，大幅提升當中的速度、準確度以及效率，徹底改變了整個研發過程的運作模式。

與傳統資料庫或搜尋引擎的被動檢索功能不同，人工智能模型能夠主動從大型數據網絡中自行學習，模擬材料行為特質、提出假設，並改進實驗參數。這不僅讓研究人員能探索現有資訊，更可加速新材料的研制、發掘未知的規律。

楊教授的研究所運用的是高通量第一性原理計算與人工智能結合的方法，通過量子力學與機器學習為基礎的自動化模擬方法，無須實驗便能評估物料的特性。在研發下一代二維電子裝置的「高介電常數物料」中，楊教授從超過14萬種已知物質中開始探索，按帶隙與介電常數等因素作篩選，最終鎖定約1,000個具有潛力的候選材料，再透過半自動化模擬進一步篩選至約20種高性能介電材料，整個流程比傳統方法快約4倍。

楊教授研究的創新之一，是將物理知識嵌入到人工智能模型，讓模型以物理學思維去學習和分析資訊，提升模型的準確性、減少對大量數據的依賴、降低能源消耗，同時提升模型的可解釋能力。團隊最近更將不同材料的短程相互作用信息編入到圖神經網絡之中，令模型在預測吸附特性或缺陷行為等複雜材料特質時更具效能。

雖然取得突破，面對龐大的數據，模型仍需更強的運算能力與演算法。不過，隨著圖形處理芯片、並行運算技術，以及代理模型與主動學習等方法不斷進步，發現新材料的速度正不斷加快。

楊教授的研究有助推動更快速、更低成本以及更可持續的材料科技的研究與發展，亦能協助香港邁向全球人工智能驅動材料科學研究的領先地位。