SASuleiman Ahmed

Resources

Geotechnics, geoenergy and machine learning.

References for my research areas, and my notes on machine-learning tools. Papers and seminars worth reading are listed at the end.

Show

Geotechnics & geoenergy

Permafrost & cold regions

Reference data, monitoring networks and design guidance for frozen ground.

Notes 5

  1. The permafrost map shows the continuous and discontinuous zones. Its description states that one-half of Canada’s land surface is underlain by permafrost, and it is published under the Open Government Licence – Canada.
  2. The Canadian permafrost monitoring network has measured permafrost for two to three decades, which gives the long records needed to calibrate models.
  3. GTN-P collects permafrost temperature (Thermal State of Permafrost) and active-layer thickness (CALM) data in one database.
  4. The Arctic Institute PDF analyses 10 to 30 year ground-temperature records from four sites and cites the Norman Wells to Zama pipeline monitoring report (GSC Open File 5331). Check which sites it covers before using it for Norman Wells.
  5. The 5th edition of the Canadian Foundation Engineering Manual (2023) adds new chapters, including one on permafrost.

OpenGeoSys

Open-source finite-element framework for coupled subsurface processes.

Notes 5

  1. OpenGeoSys (OGS) is a scientific open-source finite-element framework for simulating thermo-hydro-mechanical-chemical (THMC) processes in the subsurface.
  2. The documentation has a user guide, developer guide, tutorials, benchmarks, a feature matrix and process information.
  3. The processes overview gives examples and workflows for setting up coupled OGS processes from scratch.
  4. The developer guide walks you step by step through setting up a native development environment, which is the route to extending the code in C++.
  5. Pages under 6.5.x are tied to a release. Start from the stable address to find the current version.

Seismic response & ground improvement

Site-response software and stone-column literature.

Notes 5

  1. DEEPSOIL analyses 1D seismic response of soil columns, with equivalent-linear and non-linear approaches.
  2. According to the DEEPSOIL page, as of 31 July 2025 the program moved to Rocscience Inc. under an exclusive licence from the University of Illinois and was reintroduced as RSSeismic. DEEPSOIL is no longer available through the University of Illinois, and all DEEPSOIL files are readable in RSSeismic.
  3. The user manual guides you through the analysis in steps, from profile definition to results.
  4. The 2026 stone-column review covers mechanisms, modelling strategies and lessons from case histories.
  5. Eurocode 8, Part 5 (EN 1998-5) covers foundations, retaining structures and geotechnical aspects of seismic design. Get it from your national standards body.

Geothermal energy

Context for the Loyola Campus project and ground heat exchanger design.

Notes 4

  1. PLAN/NET ZERØ is Concordia’s decarbonisation initiative, and the Loyola Campus project is one of its parts.
  2. Exploratory geothermal drilling at Loyola Campus ran on 14 to 16 October 2025.
  3. In May 2025 Concordia, Énergir Développement Inc. and Hydro-Québec signed a collaboration agreement to explore and implement sustainable energy solutions.
  4. The ground-heat-exchanger article is a two-part trade series on modern design practice, with a modular approach that reduces system complexity and cost.

Machine learning

GitHub

Version control essentials and merge conflicts.

Notes 6

  1. Configure Git with git config, initialise a repository with git init and check its state with git status.
  2. Clone with git clone, stage files with git add ., and create and switch to a branch in one step with git checkout -b NAME.
  3. Switch between branches with git switch, push with git push <repo> <branch> and update the local repository with git pull.
  4. Ask coworkers before opening a pull request and merging your branch into main. View change history with git show.
  5. Delete a branch locally with git branch -d NAME and remotely with git push --delete origin NAME.
  6. Resolving merge conflicts is important: git merge NAME (add --abort to cancel), then accept the current, the incoming or both changes.

GNN & PyG

Graph neural networks handle multi-dimensional, relational information. They are applied to materials science, drug development and social networks, and well beyond.

Notes 6

  1. The overall PyTorch workflow: define a network class, its parameters and the forward function. Put self. before a parameter, otherwise it is not added to the parameter list.
  2. The training metric (loss) must be differentiable. The evaluation metric need not be, and scikit-learn already provides many, so there is no need to write them by hand.
  3. Precision = tp / (tp + fp). Recall = tp / (tp + fn). F1 = 2 × precision × recall / (precision + recall). All range from 0 (worst) to 1 (best).
  4. TensorboardX tracks the torch training process. sklearn.manifold reduces dimensions, and its t-SNE visualises high-dimensional data.
  5. nn.ModuleList has more freedom than nn.Sequential and supports append.
  6. Spatiotemporal GNN deals with a static structure whose features vary in time. A spatiotemporal graph captures the temporal patterns and spatial information of things’ usage events: nodes are locations, timestamps and things, and edges are relationships based on similarities and periodic patterns. The features change along time, while nodes and edges stay connected.

MPC

Model predictive control: implementation and a short introduction.

Notes 3

  1. Basic concept: the strategy u at time k is the optimal short-horizon control starting at k. It is applied to the system, which sends its updated states back to the MPC for time step k+1.
  2. Strength: it can include constraints and work with nonlinear systems (linearised if needed). Weakness: it runs online and is computationally expensive, because the future horizon is rescheduled at every step on receiving feedback.
  3. A compromise is linear parameter-varying control. With faster computing, handling the nonlinear system directly is becoming feasible.

Model-based RL

Coming soon. Notes and links will be added here.

Papers & seminars