Ph.D. Thesis Colloquium of
Ms. Anusheya M
[Research Supervisor: Dr. M. Mathirajan]
Date: 19th January 2026 [Monday]
Time: 04:30 PM
Venue: Seminar Hall [Management Studies]
Title:
“Development of Energy-Aware Intelligent and Efficient Scheduling Algorithms for Dynamic Scheduling of Diffusion Furnace(s) in Semiconductor Manufacturing under Time-Window Constraints”
Abstract:
This thesis addresses the Dynamic Scheduling (DS) of Diffusion Furnaces (DF), considering both Single Diffusion Furnace (called as DS-SDF) and Non-identical Parallel Diffusion Furnaces (called DS-NPDF) with Machine Eligibility Restrictions (MER) situations with incompatible job families along with individual jobs having characteristics such as different release time, different due-date, non-aggregable release time and due-date, time-window to be processed in DF, and time dependent electricity cost for processing at DF. The objective is to minimize total cost, comprised of tardiness cost, time-window violation cost, and electricity cost under Time of Use (TOU) pricing policy.
To gain structural insight into the problem, (0-1) mixed-integer nonlinear programming (MINLP) model for each of the situations: DS-SDF and DS-NPDF with MER is developed. Workability of the proposed MINLP model for each of the situations: DS-SDF and DS-NPDF with MER is demonstrated by developing a Small-scale numerical example and by solving using LINGO, an optimization solver. Also demonstrated the computational complexity of the proposed mathematical models by a set of small-scale problems.
Due to the computational intractability of the mathematical model for each of the situations: DS-SDF and DS-NPDF with MER, this study proposes a dispatching-rule based Greedy Heuristic Algorithm (GHA). Particularly this study proposes 50 variants of Apparent Tardiness Cost (ATC) rule-based GHA (ATC-GHA). As ATC rules are widely used in scheduling of batch processing machine in general and particularly in semiconductor manufacturing, this study has considered 10 ATC rules (seven existing and three new rules) for computing job-priority index for constructing batches and 5 batch-priority index (BPI) rules (4 existing and one new) for selecting a batch for scheduling at every decision-making time epoch. To understand the performance quality of the proposed 50 variants of ATC-GHA (a) an experimental design is proposed and 270 problem instances are generated randomly, (b) for small scale problem instances, optimal solution obtained from mathematical model is considered as benchmark solution, and for large scale problem instances generated from experimental design, estimated optimal solution (EOS) is considered as benchmark solution, and (c) multiple empirical and statistical performance measures are adopted keeping the triangulation approach. Based on comprehensive empirical and statistical performance analyses, it is observed that each of the 10 ATC rules considered in each of the proposed variants are having impact on the solution quality and each of the 5 BPI rule considered along with each of the ATC rule is not having impact on the solution quality of the proposed variants of ATC -GHA. This could be due to the fact that batch cost component inherently governs the selection of the next batch, thereby diminishing the marginal impact of the BPI rule. From these observations, it is finally concluded that (a) for DS-SDF, the proposed variant: ATC-GHA which uses the ATC rule: ATC-E (which is the proposed ATC rule in this study) is the top performing variant, and (b) for DS-NPDF with MER, the proposed variant: ATC-GHA which uses the ATC rule: ATC-Vimala is the top performing variant.
For further enhancing the solution quality obtained from the proposed variants of ATC-GHA for each of the 270 randomly generated problem instances, the metaheuristic algorithms – Simulated Annealing (SA), Tabu Search (TS), Ant Colony Optimization (ACO), and Cuckoo Search (CS) are applied with parameters of each of the metaheuristic algorithms tuned by applying the Taguchi method. Considering 50 feasible solutions obtained using each of the 50 proposed variants of ATC-GHA and the 4 feasible solutions obtained using each of the 4 metaheuristic algorithms w.r.t. each of the 270 problem instances, the EOS is obtained for performance evaluation of each of the metaheuristic algorithms. Accordingly based on both empirical and statistical performance analyses, it is observed that the metaheuristic algorithm: CS consistently achieving the best performance for both DS-SDF and DS-NPDF with MER.
Finally, this study introduces a Deep Reinforcement Learning (DRL) approach based on Proximal Policy Optimization (PPO) integrated with a Graph Neural Network (PPO-GNN) for dynamic scheduling of the diffusion furnaces. To evaluate the performance quality of the PPO-GNN approach, the same set of 270 problem instances considered in this study is employed. For each of the 270 problem instances the revised EOS are obtained using the results from PPO-GNN along with the solutions obtained from 50 proposed variants of ATC-GHA and 4 metaheuristic algorithms: SA, TS, ACO and CS. Based on the empirical and statistical performance analyses, it has been observed that for DS-SDF the metaheuristic algorithm: CS has outperformed PPO-GNN approach whereas for DS-NPDF with MER, the PPO-GNN approach has outperformed each of the four metaheuristic algorithms considered in this study. Overall, PPO-GNN provides an alternative scheduling approach, for dynamic scheduling of diffusion furnaces with MER.
Overall, this study proposed efficient composite dispatching rule based greedy heuristic algorithms to metaheuristic algorithms to AI based approach for Energy-Aware Dynamic Scheduling of Diffusion Furnace(s) in Semiconductor Manufacturing under Time-Window Constraints with cost optimization.
ALL ARE WELCOME