Update research projects

research/fire.md

@@ -11,8 +11,7 @@ banner_color: style5
 # Background
 Modern residential construction has shifted toward highly airtight homes for energy efficiency. This creates oxygen-limited conditions when a fire occurs. Unlike the well-ventilated fires that historical fire codes and research were built around, these under-ventilated fires burn slower and produce significantly more toxic combustion byproducts, posing new hazards that existing tools are not designed to handle. A key challenge in studying these fires is that the most informative physical signal, the mass loss rate (MLR), whose peak marks the precise transition from well-ventilated to under-ventilated combustion, cannot be measured in a real fire. While gas sensors (CO, O₂, NO₂, NH₃, CH₄, VOCs) can be deployed in practice, regime transitions are difficult to identify directly from raw sensor signals because the relevant combustion behavior is embedded within a complex high-dimensional sensor space. 
 
-<!-- # Stefan research -->
-<!-- <img src="../assets/images/fire.jpg" style="max-width:900px;width:100%"> -->
+# A Framework for High-Dimensional Fire Sensor Data Analysis
 
 Large-scale fire experiments generate hundreds of simultaneous sensor measurements that are difficult to interpret using conventional analysis. We developed a <b>three-step framework (Initial Screening and Visualization, Time Segmentation, and Targeted Sensor Correlation Analysis)</b> to extract physically meaningful structure from this high-dimensional data. We used <b>PCA</b> to compress 10 sensor measurements into two dimensions retaining 89% of total variance and identified a faulty CO₂ sensor that had exceeded its operational range. Next, <b>k-means clustering</b> identified three distinct fire regimes with the boundaries aligning precisely with the well-ventilated to under-ventilated transition. <b>Sparse regression models</b> were leveraged within each regime to achieve R² values of 0.96, 0.90, and 0.91, compared to 0.59 for a single global model. This demonstrates that accounting for fire regime structure is essential for accurate and interpretable sensor-based characterization.
 

research/surrogates.md

@@ -11,8 +11,13 @@ banner_color: style5
 # Background
 Engineering processes and systems are shaped by nonlinear dynamics, physical constraints, and phenomena that are difficult to describe fully using first principles. To address these modelling gaps, hybrid physics models and machine learning have emerged as viable methods for predicting system behaviour. However, classic neural networks frequently deviate from underlying physical laws, yielding incorrect or infeasible predictions. These physical inconsistencies make standard models non-ideal for constrained optimization and <b>nonlinear model predictive control (NMPC)</b> frameworks. Our research addresses these issues by developing hard-constrained and physics-guided machine learning methods that enforce physical laws and system constraints during training. Specifically, we focus on <b>hard-constrained physics-informed neural networks</b>, <b>hybrid neural differential algebraic equations</b>, and optimization-based training tailored for control applications.
 
+# Hard Constrained Neural Networks for PSE Applications
+<img src="../assets/images/pl_kkt_hpinns.png" style="max-width:700px;width:100%">
+
+<b>Physics-informed neural networks (PINNs)</b> are powerful tools for data-driven modeling of complex physical systems. However, because physical equations are typically included only as soft penalties in the training loss, PINNs do not guarantee constraint satisfaction during prediction. This has motivated the development of <b>hard-constrained neural network surrogates</b> that enforce physical laws by construction. Our research focuses on developing non-iterative, computationally efficient KKT-based frameworks for chemical process modeling, aiming to build accurate, physically consistent surrogate models for optimization and control.
+
 # Optimal Control with Neural Differential Algebraic Equations
-<img src="../assets/images/neural_mpc.png" style="max-width:900px;width:100%">
+<img src="../assets/images/neural_mpc.png" style="max-width:700px;width:100%">
 
 Recent advances in training <b>neural differential algebraic equations</b> (Neural-DAE) allow model parameters to be optimized simultaneously with hard constraints. Yet, these models have not been applied to nonlinear optimal control. Our research implements hard-constrained Neural-DAEs within <b>nonlinear model predictive control</b> frameworks using [InfiniteOpt.jl](https://infiniteopt.github.io/InfiniteOpt.jl/stable/) and GPU-acceleration via [InfiniteExaModels.jl](https://github.com/infiniteopt/InfiniteExaModels.jl) to provide accurate, scalable control of nonlinear dynamic systems.