TL;DR Overview

The Sci-ImageMiner competition defines four complementary tasks that evaluate multimodal understanding of scientific figures in Atomic Layer Deposition and Etching (ALD/E) research. Task 1 targets chart-type Classification, Task 2 focuses on Data Table Extraction from plots, and Task 3 assesses Summarization of key scientific insights conveyed by figures. Task 4 introduces Visual Question Answering across four scientific reasoning categories, requiring systems to interpret complex visual and textual elements. Together, these tasks provide a comprehensive benchmark for visual scientific reasoning in a highly specialized domain. 

Task Description

For the competition, we define four complementary image-comprehension tasks that will be grounded in ALD/E scientific figures. The four tasks are briefly described as follows:

Task 1: Classification

A supervised multi-class image classification task. Systems must identify the predefined 47 figure classes 

• Inputs
  A scientific figure extracted from an ALD/E research paper.
  Optional metadata, such as figure caption text, may be provided as context.

• Output

Predicted class labels corresponding to one of 23 predefined chart types.

Task 2: Data Table Extraction

This task focuses on structured reconstruction of the underlying tabular data encoded in scientific charts. Systems must identify column/field labels, table structure, and the textual or numerical values in each cell, producing a machine-readable Markdown representation of the data shown in the quantitative chart.

• Input

A scientific chart or plot image (e.g., bar chart, line chart, scatter plot, spectra chart, phase diagram).

• Output

A Markdown-formatted table containing:

• • Field/column names of the visualized entities

  • Extracted textual and/or numeric cell values

Task 3: Summarization

The goal of this task is to generate concise, factual summaries that capture the key trends, relationships, and scientific insights presented in the figure. Systems must demonstrate accurate semantic interpretation grounded in the visual content, possibly supported by the caption when available.

• Input

  • A scientific chart or plot image

  • figure caption text (optional).

• Output

A short textual summary (1–3 sentences) describing the main trends and takeaways.

Task 4: Visual Question Answering (VQA)

This task tests fine-grained reasoning over scientific figures by requiring systems to answer natural-language questions that reference the visual content, including axes, legends, and data patterns. The VQA task is divided into four scientifically meaningful sub-tasks:

Sub-task 4.1: Process-Oriented

Assesses understanding of ALD/E cycles, precursor chemistry, and reaction mechanisms.

Sub-task 4.2: Comparative/Trend

Evaluates reasoning about how experimental variables (temperature, cycle count, pulse length, etc.) influence outcomes such as growth rate or film thickness.

Sub-task 4.3: Structure-Property

Tests the ability to link precursor families or chemical structures to material properties such as thermal stability or growth characteristics etc.

Sub-task 4.4: Application/Performance

Measures reasoning related to device-relevant outcomes such as luminescence behavior or photovoltaic performance etc.

• Inputs

  • A scientific chart or plot image

  • A natural-language question (e.g., "At what temperature does the maximum growth rate occur?")
    
    

• Outputs

  • Yes/No: "yes" or "no"

  • Factoid: a textual term (e.g., "O₂ plasma")

  • List: comma-separated values (order-insensitive)

  • Paragraph: ≥ 3 sentences providing an explanatory answer