Machine learning is revolutionizing fundamental science by tackling long-standing mathematical challenges. A key example is the classification of topological phases of matter. While topological invariants have been essential for characterizing these phases, no single invariant works universally. This limitation has led to many phases, once considered trivial, being reclassified as topological over the years. The recent discovery of non-Hermitian band topology has intensified efforts to classify these new phases, yet existing invariants still fail to capture all their unique features. Now, researchers from Tongji University, the Chinese university of Hong Kong and Nanyang Technological University have developed a machine-learning algorithm that performs unsupervised classification of symmetry-protected non-Hermitian topological phases—without relying on any predefined topological invariants. The method can autonomously construct a topological periodic table, bypassing the need for advanced mathematics. The learning process also yielded a formula revealing how parity transformation affects periodicity. Additionally, the algorithm accounts for boundary effects, allowing investigation of how open boundaries influence the topological phase diagram. These findings establish a powerful unsupervised framework for identifying non-Hermitian topological phases, uncover previously hidden topological traits, and offer valuable insights for future theoretical and experimental work.

According to a new meta-analysis, acute-on-chronic liver failure (ACLF) can develop even in patients with previously compensated liver disease and is associated with a high risk of death. Reviewing studies published over the past decade, researchers found that short-term mortality remains high, despite shifts in underlying causes, with alcohol-related liver disease increasingly replacing viral hepatitis. The findings underscore the need for stronger prevention efforts and improved clinical management.

Scientists have developed a new computational technique to control the direction and interference of light beams without using any physical modulators. This method simplifies optical systems and could impact fields like astronomy and biomedical imaging.

Major Depressive Disorder (MDD) is a leading cause of disability among adolescents, yet the efficacy of Electroconvulsive Therapy (ECT)—a gold-standard treatment for severe cases—varies significantly among individuals. Predicting who will respond to treatment remains a challenge due to the complex interplay of brain structure and function. To address this, a research team led by Dr. Du Lei from Chongqing Medical University developed a novel deep learning model, the Dual-Branch Graph Attention Network (DBGAN). By fusing multi-modal MRI data, this model successfully predicts ECT treatment response with high accuracy, paving the way for personalized treatment strategies in adolescent psychiatry.