Welcome to the Production and Operations Analytics (POA) Research Laboratory in the Department of Mechanical, Automotive and Materials Engineering at the University of Windsor. The laboratory was previously the Production and Operations Management Research Lab.
A message from the lab director
The planning problems that matter on a factory floor do not stay still. Part geometry changes, the product mix shifts, and a disruption on a machine, a vehicle, or in a clinic can undo a plan that was optimal an hour earlier. This lab builds the planning functions that keep those decisions workable: computer-aided process planning, scheduling, lot sizing, and layout, joined to machine learning and carried into operation by digital twins.
We treat sustainability, flexibility, and predictability as part of the model. The same methods now support hybrid manufacturing, operating-room planning, and the siting of electric-vehicle charging and distributed generation.
If you are a student looking for a thesis, start with the Teaching page and the People page. If you are a company with a planning problem that does not fit a textbook, the industry projects page shows the kinds of problems we have already taken on. Write to azab@uwindsor.ca.
Dr. Ahmed Azab, P.Eng.
Professor and Lab Director
Research program
The program develops the planning functions of smart manufacturing. The work sits at the junction of operations research and machine learning. The aim is process plans and shop-floor decisions that stay computationally workable as part geometry, product mix, and operating conditions change. Computer-aided process planning and production scheduling are the core. Digital twins carry those plans into real-time operation.
Computer-aided process planning
Macro-level process planning has been part of this program since 2007. The current line, CAPP-GPT, trains generative transformer architectures on data-enriched, optimized plans. The planners sequence hybrid-manufacturing features and choose orientation, setup, and tooling, taking account of productive and non-productive time, precedence, and technological restraints.
Published models follow a single processing strategy. The lab is extending them so that part geometry can be decomposed in more than one way, recognized from boundary-representation and constructive-solid-geometry data, and labeled additive or subtractive once the stock envelope is known. Convolutional sampling of NURBS faces and edges, together with a graph transformer, encodes the CAD model. A decoder writes the operation sequence. A second, contrastive architecture places geometry and manufacturing labels in one embedding, so that new feature classes can be added through the label side.
The corpus is built by exact sequencing and setup grouping on public CAD sets and on parts generated in NX Open, with logical analysis of data carrying the larger instances, and by clustering part families so that a later part customizes a master plan. Stock design is solved as constrained clustering into cuboid billets. The same sequencing and clustering methods support platform design for hybrid manufacturing and smart assembly on transfer lines. The intent is an independent module for existing product-lifecycle systems: shorter planning time, higher plan quality, and lower cost and energy.
Process parameters
Manufacturing process characterization is the micro-planning line. A mixed-integer program relates surface roughness to the machining parameters through Taguchi’s quadratic loss function, and it accounts for failure rate, greenhouse gases, energy consumption, and lot size.
For additive manufacturing on a five-axis setup, in-situ monitoring and statistical analysis characterize multi-track height. Process parameters, including shield-gas flow rate, affect melt-pool temperature and digital imagery, and those signatures are used to predict the process response. That opens a path to in-situ monitoring and self-healing, which reduces yield loss and energy consumption.
Scheduling and digital twins
Scheduling covers smart, flexible manufacturing and healthcare delivery. For manufacturing, the lab uses a two-stage digital twin.
The digital-model stage produces fairness-aware nominal schedules that minimize makespan and a quadratic, weight-proportional tardiness penalty, so that any shared tardiness is split more equitably across customers. Smaller instances are solved to certified optimality. Buffer sizes and smart automated guided vehicles are planned in the same stage. A deep network estimates material-handling time so that an evolutionary algorithm can balance makespan and work-in-process inventory.
Once the system is running, the twin ingests machine health, vehicle position, job status, and buffer level, and it tests corrective actions before they are released to the floor. A newly arrived job is treated as an insertion. A job-level trigger identifies which existing operations the arrival can displace, and an exact model re-optimizes only that scope. On standard flexible-job-shop benchmarks, the scope-restricted plans reach a mean solution-quality gap of 1.5% relative to complete rescheduling, compared with 60.7% for right-shift, while the search space stays restricted.
The same scheduling line includes reconfigurable manufacturing systems and assembly lines with moving labour and a limited workspace. In healthcare, robust optimization addresses stochastic service times, no-shows, and appointment mode. Stochastic multi-objective models schedule operating rooms across surgeons, rooms, recovery beds, and patients. Those algorithms sit inside a decision-support system that spans strategic, tactical, and operational planning.
Lot sizing and delayed product differentiation
These models decide how much to make, whether a setup carries into the next period, and when product variety should be postponed. They already include economic lot sizing, setup carryover, and greenhouse gases. A hybridized Dantzig-Wolfe and dynamic programming algorithm solves the multi-level capacitated lot-sizing problem. Setup carryover is modeled as a maximum-weight independent set in a chain of cliques. The same line extends a multi-level product platform to batch production that uses hybrid manufacturing and smart automated guided vehicles.
Layout and supply chains
Dynamic and robust facility layout is treated by hybridizing dynamic programming with heuristics. Work-in-process is rerouted when the product mix changes. A semi-robust formulation for cellular manufacturing allows only the pickup and drop-off points to change across periods. These layouts shorten part routes and thereby cut cost and energy use.
Beyond the factory, machine learning is used to quantify risk in the design of distributed supply chains, and those predictions enter integer linear programs for facility location and allocation.
Power grids and healthcare
The same mathematical models and hybrid algorithms are being extended to other sectors.
For electric vehicles, the lab models the siting and sizing of distributed generation and charging infrastructure. To mitigate the intermittency of renewable generation and to regulate voltage, routing and scheduling models deploy mobile energy storage and mobile STATCOMs where the grid needs them.
For healthcare, the lab has developed a decision-support system for operating-room planning under uncertain arrivals and no-shows, using robust optimization, hyper-heuristics, and discrete-event simulation.