Machine learning (ML) has become a powerful tool in molecular simulations, driving progress across multiple fronts [1]. A major breakthrough has been the development of machine learning interatomic potentials (MLIPs), which offer near ab initio accuracy at much lower computational cost, enabling large-scale and long-time molecular dynamics simulations [2].

At coarser resolutions, ML has enabled the construction of effective force fields for coarse-grained models, trained on high-resolution data to capture essential physics while accessing longer timescales [3]. In parallel, generative models are being applied to learn conformational ensembles directly, offering new strategies for sampling complex molecular landscapes without explicit trajectory generation [4,5].

These developments have greatly expanded the scope of molecular modeling—but they also raise several fundamental questions that must be addressed to realize the full potential of ML in this domain:

• Transferability: How can models generalize across different molecular systems and conditions, especially when data is scarce?
  
  
• Reliability: How do we quantify uncertainty and ensure the robustness of ML-driven simulations?
  
  
• Scalability: Can ML models be applied effectively to larger systems or longer time scales without compromising performance?
  
  
• Bridging timescales: How can we connect short-time dynamics captured by MLIPs with long-time conformational sampling and extract kinetic observables (e.g., rates, pathways) from ML models?
  

This workshop will bring together researchers from across computational chemistry, physics, applied maths, and ML to discuss recent advances and identify the most pressing challenges at these intersections. By fostering dialogue across disciplines, it aims to both disseminate recent methodological progress and guide the development of next-generation approaches in data-driven molecular science.