Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CLIFF: Continual Learning for Incremental Flake Features in 2D Material Identification

Sankalp Pandey, Xuan Bac Nguyen, Nicholas Borys, Hugh Churchill, Khoa Luu

TL;DR

This paper proposes a new Continual-Learning Framework for Flake Layer Classification (CLIFF), the first systematic study of continual learning in the domain of two-dimensional materials, and achieves competitive accuracy with significantly lower forgetting than naive fine-tuning and a prompt-based baseline.

Abstract

Identifying quantum flakes is crucial for scalable quantum hardware; however, automated layer classification from optical microscopy remains challenging due to substantial appearance shifts across different materials. This paper proposes a new Continual-Learning Framework for Flake Layer Classification (CLIFF). To the best of our knowledge, this work represents the first systematic study of continual learning in two-dimensional (2D) materials. The proposed framework enables the model to distinguish materials and their physical and optical properties by freezing the backbone and base head, which are trained on a reference material. For each new material, it learns a material-specific prompt, embedding, and a delta head. A prompt pool and a cosine-similarity gate modulate features and compute material-specific corrections. Additionally, memory replay with knowledge distillation is incorporated. CLIFF achieves competitive accuracy with significantly lower forgetting than naive fine-tuning and a prompt-based baseline.

CLIFF: Continual Learning for Incremental Flake Features in 2D Material Identification

TL;DR

This paper proposes a new Continual-Learning Framework for Flake Layer Classification (CLIFF), the first systematic study of continual learning in the domain of two-dimensional materials, and achieves competitive accuracy with significantly lower forgetting than naive fine-tuning and a prompt-based baseline.

Abstract

Identifying quantum flakes is crucial for scalable quantum hardware; however, automated layer classification from optical microscopy remains challenging due to substantial appearance shifts across different materials. This paper proposes a new Continual-Learning Framework for Flake Layer Classification (CLIFF). To the best of our knowledge, this work represents the first systematic study of continual learning in two-dimensional (2D) materials. The proposed framework enables the model to distinguish materials and their physical and optical properties by freezing the backbone and base head, which are trained on a reference material. For each new material, it learns a material-specific prompt, embedding, and a delta head. A prompt pool and a cosine-similarity gate modulate features and compute material-specific corrections. Additionally, memory replay with knowledge distillation is incorporated. CLIFF achieves competitive accuracy with significantly lower forgetting than naive fine-tuning and a prompt-based baseline.

Paper Structure

This paper contains 26 sections, 8 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. The Challenges of Automated Flake Identification
  3. Limitations of Prior Work
  4. Problem Motivation
  5. Contributions of this Work
  6. Related Work
  7. Automatic 2D Materials Thickness Estimation
  8. Continual Learning
  9. The Proposed Method
  10. Base Training on a Reference Material
  11. Incremental Learning for New Materials
  12. Prompting
  13. CLIFF Head
  14. Optimization with Rehearsal and Distillation
  15. Complexity
  16. ...and 11 more sections

Figures (4)

  • Figure 1: Comparison of 2D material flake characterization methods. (a) Automated thickness identification using deep learning on optical microscopy images offers a fast, scalable approach capable of processing large numbers of flakes and predicting their layer count. (b) The traditional approach relies on Atomic Force Microscopy (AFM), which is highly accurate but time-consuming and unscalable for high-throughput discovery.
  • Figure 2: The proposed CLIFF approach.
  • Figure 3: Internal architecture of the CLIFF Head. This illustrates the pipeline for new-image features. Replayed features follow the same process to produce student prediction scores.
  • Figure 4: Visual comparison of classification predictions on a Thick_BN flake (Task 1) after training on all four tasks. The bounding box indicates the input crop. CLIFF (Green) correctly predicts the label, whereas Naive Fine-tuning and L2P (Red) fail due to catastrophic forgetting.