|
|
|
||
|
This course introduces students to the intersection of Artificial Intelligence (AI) and Computational Chemistry. It is
designed for students with a background in chemistry but assumes no prior experience in programming or machine learning. The course takes a practical, ”hands-on” approach. Students will journey from basic AI concepts (teaching a computer to recognize handwritten digits) to advanced scientific applications (training a neural network to accurately reproduce the anomalous physical properties of liquid water from quantum mechanical first principles). By the end of the semester, students will have built two distinct computational models: a classical physical model and a state-of-the-art Neural Network Potential (NNP). Last update: Ušelová Kateřina, RNDr., Ph.D. (29.04.2026)
|
|
||
|
Artificial Intelligence in Computational Chemistry
Syllabus – One Semester (12 Weeks)
Course Overview This course introduces students to the intersection of Artificial Intelligence (AI) and Computational Chemistry. It is designed for students with a background in chemistry but assumes no prior experience in programming or machine learning.
The course takes a practical, ”hands-on” approach. Students will journey from basic AI concepts (teaching a computer to recognize handwritten digits) to advanced scientific applications (training a neural network to accurately reproduce the anomalous physical properties of liquid water from quantum mechanical first principles). By the end of the semester, students will have built two distinct computational models: a classical physical model and a state-of-the-art Neural Network Potential (NNP).
Learning Objectives By the end of this course, students will be able to: 1. Demystify AI: Understand Machine Learning as a tool for function approximation and optimization. 2. Implement ML Pipelines: Use Python and PyTorch to build, train, and validate simple neural networks. 3. Bridge Theory and Practice: Connect quantum mechanical data (Psi4) to classical molecular dynamics (OpenMM). 4. Utilize HPC Resources: Gain experience submitting jobs to a computational grid/cluster. 5. Evaluate Models: Critically compare ”Physical” models (interpretable but rigid) vs. ”Neural” models (accurate but opaque).
Software Stack All practicals will use open-source, Python-based tools: • Language: Python (Jupyter Notebooks) • AI/ML: PyTorch (TorchANI / NequIP) • Quantum Chemistry: Psi4 • Molecular Dynamics: OpenMM • Visualization: Matplotlib, py3Dmol • Trajectory Analysis: MDTraj
Weekly Schedule Block 1: Foundations of AI – ”From Pixels to Logic” (Weeks 1–3) Goal: Build confidence with Python and the core concepts of Machine Learning using visual data.
Topic: AI is just ”fancy curve fitting.” Activity: Python crash course. Generating noisy data for simple functions (y = sin(x)) and using PyTorch to optimize polynomial coefficients to fit the data. Introduction to Loss Functions and Gradient Descent.
Topic: Neural Networks, Pattern Recognition, and Reversibility. Activity: Two inverse tasks on MNIST: (1) Recognise – Classifier (pixels → label); (2) Draw – given a label, generate the digit image, solved two ways: Autoencoder and GAN. Key Concept: Recognition and drawing are inverse problems – the same goal, two very different strategies.
Topic: Transformers and sequence-based models. Activity: Conceptual overview of the Transformer (attention, next-token prediction). Hands-on: use MolGPT to generate novel molecules from SMILES. Introduction to ChemCrow as an LLM chemistry agent. Key Concept: SMILES strings are a language for molecules – the same machinery that powers ChatGPT applies to chemical space.
Block 2: Project 1 – The Classical ”White Box” Model (Weeks 4–7) Goal: Train a physically interpretable model (Lennard-Jones) on quantum data.
Topic: The Potential Energy Surface (PES). Activity: Using Psi4 to calculate the interaction energy of a water dimer at various distances. Deliverable: A dataset of Distance (r) vs. Energy (E).
Topic: Parameter Fitting. Activity: Defining the Lennard-Jones potential in PyTorch. Using the optimizer to find the best σ and ε parameters that match the Week 4 quantum data.
Topic: From Parameters to Motion. Activity: Injecting the learned parameters into OpenMM. Running a simulation of 500 water molecules.
Topic: Did we create water? Activity: RDF and density accumulated on-the-fly during the simulation. Visualising the molecular motion with py3Dmol. Outcome: Observing the limitations of simple physical models (e.g., incorrect density or phase).
Block 3: Project 2 – The Neural ”Black Box” Model (Weeks 8–11) Goal: Train a Neural Network Potential (NNP) to learn the physics from scratch.
Topic: How NNPs work and what is already available. Activity: Introduction to the architecture of NNPs (symmetry functions, equivariance). Loading the pre-trained ANI-2x model via TorchANI and evaluating it on water configurations from Week 4. Comparing ANI-2x energies against the Psi4 reference data. Key Concept: A pre-trained NNP is the reliable baseline for the MD comparison in Week 10. Training from scratch is a research-level task – we will attempt it, but without guaranteed success.
Topic: Scaling up data generation. Activity: Introduction to the Computational Grid. Submitting batch scripts to generate a dataset of distorted water clusters using Psi4.
Topic: The ”Brain” driving the Simulation. Activity: Interfacing ANI-2x with OpenMM (openmm-torch) and running the MD simulation. Optionally: attempt to train NequIP on the Week 9 dataset and substitute the student model if stable. Concept: Overfitting vs. Generalization in high-dimensional space.
Topic: The Showdown. Activity: Visualising both simulations with py3Dmol. Plotting the on-the-fly RDFs of Project 1 (Classical), Project 2 (Neural), and Experimental Data on a single graph.
Block 4: Conclusion (Week 12)
Activity: Student presentations summarizing their workflow. Discussion on the future of AI in chemistry (AlphaFold, autonomous labs).
Assessment Methods • Weekly Labs (40%): Completion of Jupyter Notebooks. • Midterm Project (30%): The Classical Water Model (Weeks 4-7). • Final Project (30%): The Neural Water Model & Final Presentation (Weeks 8-12). Last update: Ušelová Kateřina, RNDr., Ph.D. (29.04.2026)
|
| The course does not include work placement |