MP-GFormer: A 3D-Geometry-Aware Dynamic Graph Transformer Approach for Machining Process Planning
Fatemeh Elhambakhsh, Gaurav Ameta, Aditi Roy, Hyunwoong Ko
TL;DR
This paper introduces MP-GFormer, a novel 3D-geometry-aware dynamic graph transformer that effectively models evolving part geometries to improve machining process planning accuracy.
Contribution
It presents a new transformer-based approach that incorporates 3D geometric information into dynamic graph learning for machining process sequence prediction.
Findings
Achieves 24% accuracy improvement in main operation prediction
Achieves 36% accuracy improvement in sub-operation prediction
Demonstrates effectiveness on a synthesized dataset
Abstract
Machining process planning (MP) is inherently complex due to structural and geometrical dependencies among part features and machining operations. A key challenge lies in capturing dynamic interdependencies that evolve with distinct part geometries as operations are performed. Machine learning has been applied to address challenges in MP, such as operation selection and machining sequence prediction. Dynamic graph learning (DGL) has been widely used to model dynamic systems, thanks to its ability to integrate spatio-temporal relationships. However, in MP, while existing DGL approaches can capture these dependencies, they fail to incorporate three-dimensional (3D) geometric information of parts and thus lack domain awareness in predicting machining operation sequences. To address this limitation, we propose MP-GFormer, a 3D-geometry-aware dynamic graph transformer that integrates…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
Technical Innovation: Filling the Integration Gap between 3D Geometry and Dynamic Graph LearningExisting DGL methods in MP only focus on sequential graph structures, ignoring the core domain knowledge of part 3D geometry, which leads to models lacking domain awareness. MP-GFormer is the first to deeply integrate 3D geometric information into a dynamic graph Transformer framework: it captures the post-machining geometric evolution through process graph sequences converted from STL files, obtains
Insufficient Model Transferability: Dependence on "Full-Operation Geometric Sequence" Training Limits Cross-Scenario ApplicationThe training of MP-GFormer in the paper relies on "the sequence of STL geometric graphs of the part after each machining operation"—that is, complete geometric evolution data of the part from raw material to finished product (geometric state corresponding to each operation) is required to complete model training. However, in practical industrial scenarios, there are sig
The paper demonstrates strong originality through its novel formulation of machining process planning as a geometry-aware dynamic graph learning problem. While dynamic graph learning has been applied to various domains, existing approaches fail to capture the essential domain constraint that part geometry fundamentally determines feasible machining operations. The key innovation lies in the creative integration of two complementary geometric representations: STL-based process graphs capturing ev
The experimental validation relies entirely on synthesized data generated from only 6 parameterized base geometries (Figure 3) with 2.5-axis planar machining, severely limiting generalizability claims. Real manufacturing involves diverse multi-axis operations, material constraints, tool wear, and fixture considerations absent from this evaluation. The paper provides no validation on actual industrial process plans or analysis of synthetic-to-real transfer challenges. With only 3 main and 12 sub-
The paper cleanly scopes machining process planning to macro-level sequencing and builds a geometry-aware model that fuses the initial B-Rep design with per-step STL geometry, using GAT encoders for local structure, cross-attention to inject design intent, and a causally masked Transformer decoder to respect step order. It defines concrete node/edge features (centroids, normals, dihedral angles, surface types, etc.), reports strong gains over dynamic-graph baselines on a standardized synthetic d
The method’s output is non-executable: it predicts only main/sub operation labels per step without binding them to concrete geometric targets or proposing tools, parameters, or toolpaths, so it cannot yield a CAM-ready plan. Empirical validation relies on a small, synthetic, parameterized dataset with no evidence of transfer to real shops, machines, fixtures, or tool libraries, which limits external validity. The graph granularity is misaligned with machining semantics—modelling each STL triangl
- The paper is generally well-written and easy to read. The motivation for the approach is clear and convincing. - The paper introduces a novel and relevant formulation of machining process planning as a dynamic graph learning problem conditioned on some target design. This approach of modeling the sequence of manufacturing steps as an evolving series of graphs is an interesting and appropriate way to capture the problem's inherent temporal and geometric dependencies. - The work addresses an imp
- The proposed architecture, MP-GFormer, is presented as a complex sequence of components (GAT encoders, cross-attention, transformer decoder, etc.) without sufficient justification for each design choice. The paper does not explain why this specific, convoluted combination of modules is necessary over simpler alternatives, such as a temporal GNN. - The experimental evaluation does not convincingly support the paper's claims. The hyperparameter tuning and ablation studies do not sufficiently con
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsManufacturing Process and Optimization · 3D Shape Modeling and Analysis · Robot Manipulation and Learning