Zhongyu Zhang

DCS Laboratory Development and Industrial Training

May 2026 - Present / SUPCON Technology

• Completed hands-on training led by SUPCON engineers and participated in on-site commissioning of the ECS-700 industrial distributed control system and coupled-tank level-control hardware.
• Collaborating with the laboratory team to design course experiments that teach DCS fundamentals, engineering software configuration, and closed-loop process control using the industrial platform.

Student Researcher

Feb. 2024 - Jun. 2026 / Teck Resources Limited

KIVCET Furnace Feed Optimization (Jan. 2025 - Jun. 2026)

• Collaborated with senior process control engineers at Trail Smelter to develop a Databricks-based decision-support pipeline combining plant data, machine-learning estimates of residue composition, and optimization to recommend furnace feed adjustments.
• Translated operational requirements into an uncertainty-aware optimization model balancing residue utilization, feed moisture, chemistry targets, and equipment constraints.
• Simulations improved feed element-ratio target compliance by 63% and increased monthly residue utilization by several hundred tonnes versus manual operation.

MPC Model Health Monitoring (Feb. 2024 - Jan. 2025)

• Worked with Teck engineers to migrate a model-plant mismatch monitoring toolbox from a local application to Databricks and integrate it into daily reporting for two industrial MPC systems.
• Combined five complementary analytics methods to detect when controller models no longer matched plant behaviour and identify the affected input-output relationships.
• Automated data processing and interactive HTML reporting to present operating KPIs, process trends, and model-health alerts in one place for engineering review.

Robust Machine Learning for Corrupted and Underrepresented Data

2025 - Present

• Developed two complementary robust-learning methods that either filter unreliable samples or emphasize rare but valuable operating conditions using partial optimal transport and distributionally robust optimization.
• For corrupted measurements and labels, automatically excluded unreliable samples, improving process fault-classification accuracy from 57% to 77% and reducing soft-sensor RMSE by 57%.
• For underrepresented or shifted operating regimes, emphasized difficult but informative samples, reducing degraded-regime prediction RMSE by 37% and improving sensor-drift classification accuracy from 68% to 73%.

Data-Driven Petroleum Composition Inference from Routine Measurements

2024 - 2026

• Co-developed data-driven methods to infer detailed molecular compositions of petroleum streams from limited, routinely available laboratory measurements, reducing reliance on costly specialized analysis.
• Combined neural networks, database matching, and property-constrained optimization to reconstruct chemically realistic compositions for both naphtha and heavy gas-oil fractions.
• Achieved an average R2 of 0.957 across 44 naphtha test samples; on 21 heavy gas-oil test samples, the inferred molecular patterns closely matched high-resolution measurements while reproducing key bulk properties near routine measurement accuracy.

Quantifying Process Tolerance to Uncertainty with Neural Networks

2023 - 2025

• Developed neural-network methods to map constraint-satisfying operating regions and quantify how much variation in process conditions a system can tolerate without violating operating limits.
• Extended the approach from steady-state systems to time-varying processes, using physics-informed surrogate models and classifiers to handle nonlinear, disconnected, and high-dimensional operating regions.
• Achieved up to 27x faster evaluation than conventional solver-based methods in benchmark process case studies, while maintaining over 95% accuracy in classifying feasible operating trajectories.

Physics-Informed Modeling and Dose Optimization for Renal Anemia Treatment

2022 - 2024

• Developed a physics-informed neural network that combines sparse clinical data with physiological equations to estimate patient-specific parameters and predict haemoglobin response to EPO treatment; evaluated using records from 291 patients.
• Converted the personalized model into a fast control-oriented surrogate and integrated it with zone model predictive control to optimize EPO dose schedules while maintaining haemoglobin within a target range.
• Reduced prediction RMSE by 44-48% compared with LSTM and restored haemoglobin to the target range following simulated bleeding and infection disturbances in a two-year virtual-patient study.

Distributionally Robust Optimization for Gaussian Mixture Model Ambiguity Under Moment Variations

Oral presentation in American Institute of Chemical Engineers (AIChE) Annual Meeting on November 3, 2025, Boston, Massachusetts, U.S.

Decision-focused Learning Considering Solution Robustness under Data Uncertainty

Oral presentation in Canadian Chemical Engineering Conference (CSChE 2025) on October 7, 2025, Montreal, Quebec, Canada.

Dynamic Process Operational Flexibility Analysis Using Neural Networks

Poster presentation in Canadian Chemical Engineering Conference (CSChE 2024) on October 6-9, 2024, Toronto, Ontario, Canada.

Model Predictive Control for Renal Anemia Treatment through Physics-Informed Neural Network

Poster presentation in 12th IFAC International Symposium on Advanced Control of Chemical Processes (ADCHEM) on July 14-17, 2024, Toronto, Ontario, Canada.

A Novel Volumetric Flexibility Index Calculation Method Using Physics-Informed Neural Networks

Oral presentation in American Institute of Chemical Engineers (AIChE) Annual Meeting on November 9, 2023, Orlando, Florida, U.S.