How to become expert at thing: 1 iteratively take on concrete projects and accomplish them depth wise, learning “on demand” (ie don’t learn bottom up breadth wise) 2 teach/summarize everything you learn in your own words 3 only compare yourself to younger you, never to others

Now in Digital Discovery, @_Xu_Chen_ introduces the ACES-GNN framework, designed to simultaneously improve predictive accuracy and interpretability by integrating explanation supervision for activity cliffs (ACs) into GNN training. pubs.rsc.org/en/content/arti…

(Reposting due to wrong link posted in previous post) Our recent work about utilizing domain adversarial neural network trained on simulated data to analyze experimental spectra is now featured in the cover image of Newton, volume 1, issue 3! cell.com/newton/fulltext/S29…

Prof. Mingda Li @mit_nse just published a preview in Newton to introduce our recent paper in Newton (doi.org/10.1016/j.newton.202…) about using domain adversarial neural networks (DANN) to detect phase transition in superconducting material. cell.com/newton/fulltext/S29…

Data scarcity often limits ML applications in experimental science. We tackle this with a domain-adversarial framework that fuses abundant simulation data with scarce experimental data to detect thermodynamic phase transitions from photoemission spectroscopy. 🧵1/6

Machine learning approach reveals exciton states in large molecular assemblies Organic molecular assemblies composed of aromatic building blocks have attracted wide interest because of their tunable optical behaviors in fields like photovoltaics, bio-inspired light harvesting, and nanoelectronics. Understanding how excitations in these stacked molecules evolve and combine into collective phenomena is a longstanding challenge, as direct quantum simulations often become infeasible for large assemblies. The development of systematic, data-driven methods to predict excited-state properties has become increasingly important for designing molecular aggregates with targeted performances. Ren et al. tackled this issue using a machine learning workflow that predicts exciton Hamiltonians from data generated by a fragment-based quantum approach. By focusing on dimer units, they trained multiple neural networks that accurately reproduce both local excitations and charge-transfer states. The mean absolute error for dimer coupling terms remains under 10 meV, with coefficients of determination surpassing 0.99. Once trained, this size-transferable model reconstructs the exciton Hamiltonian of arbitrarily large aggregates by combining predictions for each dimer pair, an approach validated on trimer and tetramer test sets with out-of-sample errors around 15–30 meV for low-lying excited states. Crucially, it scales efficiently, enabling the analysis of aggregates up to 50 monomers in length without direct expensive electronic-structure calculations. The authors showed that this machine learning exciton model correctly captured the oscillator strengths of perylene aggregates and illuminated how couplings influence their optical gaps. Predictions were consistent with reference quantum calculations, yet orders of magnitude faster. Using this model, the researchers demonstrated that charge-transfer couplings significantly reduce the optical gap and increase its variability with size, in line with prior experimental observations of the pronounced red shifts in large stacks. This method opens the door to better computational tools for studying nanosized organic materials, offering design insights for applications in sensing, energy storage, and optoelectronic devices Paper: pubs.acs.org/doi/full/10.102…