MatSynth Dataset Overview

• The MatSynth Dataset is a collection of open datasets comprising high-resolution PBR materials, synthetic 1D spectroscopy data, and extensive 2D material synthesis records.
• It utilizes rigorous quality assurance methods, including CLIP-based duplicate filtering, statistical augmentation, and multi-round expert validations.
• Applications range from computer vision and machine learning benchmarking to AI-driven material synthesis planning, with significant improvements in rendering fidelity and model performance.

The term "MatSynth Dataset" refers to a family of open datasets in computational materials science, computer vision, and machine learning encompassing at least three major resources with distinct foci—physically-based rendering (PBR) of materials, synthetic spectroscopic data for ML benchmarks, and large-scale compilations of 2D material synthesis procedures. Each instantiation under the MatSynth label is characterized by its scale, domain specificity, data richness, and broad academic accessibility. The following sections provide a comprehensive technical summary of the primary MatSynth datasets as reported in the scholarly literature.

1. Major MatSynth Datasets: Definitions and Scope

Three principal datasets are commonly referred to as "MatSynth," each serving different research communities:

• MatSynth PBR Materials Dataset: A collection of 4,069 ultra-high-resolution tileable physically based rendering (PBR) materials, primarily designed for computer graphics, vision, and machine learning applications (Vecchio et al., 2024).
• MatSynth Synthetic Spectroscopy Dataset: A universal benchmark comprising ≈35,000 synthetic 1D spectra emulating common characterization techniques (XRD, NMR, Raman), generated for model validation and method development in spectroscopic ML (Schuetzke et al., 2022).
• MatSyn25 (Material Synthesis 2025) Dataset: The most expansive set, cataloging 163,240 synthesis-process records for 2D materials, extracted from 85,160 peer-reviewed articles, supporting research in AI-driven material discovery and synthesis planning (Li et al., 1 Oct 2025).

Variant	Domain	Size/Scale	Primary Use Case
PBR Materials	Graphics	4,069 materials, 3.4M+ renders	SVBRDF, generation, rendering
Synthetic Spectroscopy	Spectroscopy/ML	35,000 spectra, 500 classes	ML benchmarking, model dev.
MatSyn25	2D Materials	163,240 records	Synthesis planning, LLM training

2. Data Structure, Modalities, and Rich Metadata

2.1. MatSynth PBR Materials (Vecchio et al., 2024)

Each PBR material comprises up to eight 4K texture maps: base color, diffuse, normal (OpenGL convention), height (16-bit), roughness, metallic, specular, and optional opacity. Accompanying metadata spans source, license, tags (1,239 unique), category, creation method (procedural/photogrammetry/manual/blends), stationarity, versioning, and, if available, descriptions, authorship, and physical size. Rendered examples are provided for each material: 168 crops per material under five lighting environments, yielding 3,417,960 1K renderings.

2.2. Synthetic Spectroscopic MatSynth (Schuetzke et al., 2022)

Data are stored as NumPy arrays: $X_{train}$, $X_{test}$ (shape: $N_{spec} \times N_{pts}$), with class labels ($y_{train}$, $y_{test}$). Each class is defined by a set of Gaussian peak parameters $\{\mu_i, A_i, \sigma_i\}$, with augmentation introducing systematic parameter perturbations. A JSON configuration enumerates all class definitions and augmentation magnitudes.

2.3. MatSyn25 Synthesis Dataset (Li et al., 1 Oct 2025)

Each record is a JSON object encapsulating paper metadata, material details (chemical formula in LaTeX, type, morphology), synthesis process (name, type, objectives, detailed multistep procedures including temperature, pressure, time, solvent, precursors, catalyst, equipment), post-treatment, and explicit safety notes. Extracted chemical systems include 182,299 unique 2D materials (with 74,464 chemically distinct).

3. Dataset Generation and Quality Assurance Pipelines

3.1. MatSynth PBR Materials

The collection aggregates from open-source libraries (e.g., AmbientCG, ShareTextures), with rigorous cleaning (removal of unrealistically rendered materials), duplicate filtering via CLIP embedding ($>0.99$ similarity), and augmentation (crop/rotation, height-guided blending). Normal maps are validated by comparing supplied and $\nabla$(height)-derived normals, with Y-axis inversion applied for discrepancies.

3.2. Synthetic Spectroscopy

Datasets are synthesized by summing independent Gaussian peaks: $$S(x) = \sum_{i=1}^N A'_i \exp\left[ -\frac{(x-\mu'_i)^2}{2 {(\sigma'_i)}^2} \right]$$, with parameters stochastically perturbed per spectrum. Training samples expose greater shift/variation ranges than test samples, eliminating test/train overlap. Customization (class count, peak number, variation magnitude) is configurable by open-source Python scripts.

3.3. MatSyn25 Text-to-Structure Extraction

MatSyn25 leverages an AI pipeline: structured text is parsed from PDF via MinerU (including OCR), semantically embedded, and processed by a Qwen3-8B LLM fine-tuned with LoRA using over 5,000 hand-annotated examples. The extraction proceeds in three phases: identification of processes, step decomposition, and entity/attribute linking, with multi-round expert validation on >10,000 records and post-processing for unit normalization and outlier rejection.

4. Access Modalities, Licensing, and Interactivity

• PBR Materials: Downloadable in PNG/EXR (4K) and JSON (metadata) formats with crops/renders (1K) via https://www.gvecchio.com/matsynth. Licensing is CC0 (95%), CC-BY (clearly marked subset); no academic/commercial restrictions.
• Synthetic Spectroscopy: Provided as NumPy arrays and JSON class catalogs (OSF archive: https://osf.io/pqahd, scripts: https://github.com/jschuetzke/synthetic-spectra-generation).
• MatSyn25: JSON (primary), CSV, and SQLite, totaling ≈1.3 GB (compressed). Accessible for download, API, and interactive exploration, including retrieval and RAG-augmented Q&A, via https://matsynai.stpaper.cn/. Licensed under CC BY 4.0 (GitHub, HuggingFace mirrors).

5. Research Benchmarks and Applications

5.1. MatSynth PBR Materials

Benchmarks include SVBRDF estimation, material generation (latent diffusion, e.g., MatFuse 2023), and rendering tasks. Notable outcomes: training on MatSynth yields lower RMSE/LPIPS and higher SSIM compared to Deschaintre 2018, and FID for generative tasks improves from 239.9 (pre-MatSynth) to 89.84 (trained from scratch on MatSynth) (Vecchio et al., 2024).

5.2. Synthetic Spectroscopy

Eight 1D CNN architectures (e.g., CNN2, CNN3, VGG, ResNet, Inception) have been systematically evaluated. CNN6 achieves best test accuracy (14 ± 2 misclassifications, 99.7% accuracy); ResNet does not surpass baseline architectures (Schuetzke et al., 2022). Observed: output dimensionality reduction (≤80) is a key determinant of performance, while complex layers increase runtime without accuracy gain.

5.3. MatSyn25

Applications span LLM fine-tuning for synthesis step generation, retrosynthetic planning, benchmarking of automated protocols, synthesizability screens, Bayesian/reinforcement optimization of process parameters, and knowledge-graph mining (materials ↔ methods ↔ properties). Process parameter distributions: mean temperature $\mu=165^\circ C$, $X_{test}$0; reaction times median ≈ 3 h; hydrothermal peaks at 180–220°C, CVD at 800–1000°C; strong positive Pearson correlation ($X_{test}$1) between synthesis temperature and crystallite size for TMDs (Li et al., 1 Oct 2025).

6. Limitations, Extensions, and Future Directions

• PBR Materials: Currently restricted to stationary, tileable samples, no measured BRDFs or advanced scattering phenomena; future plans include expansion to spatially varying, anisotropic, and subsurface-scattering-capable datasets and integration with procedural graph representations.
• Synthetic Spectroscopy: Only 1D Gaussian peaks simulated; extensibility via user scripts to incorporate non-Gaussian peak shapes, explicit noise, or additional modalities.
• MatSyn25: Focused on 2D material syntheses published 2000–2025; potential to extend to beyond-2D systems and couple with high-throughput experimental platforms. A plausible implication is that similar LLM-extracted approaches could be generalized for inorganic materials synthesis beyond 2D compounds.

7. Representative Example Records

PBR Materials: Metadata Schema (Editor’s Term)

Field	Example Entry	Description
Name	walnut_dark_019	Material designation
Category	wood	Superclass label
Maps	basecolor.png, normal.png	4K tileable textures
Renderings	render_env01.png	1K crops, 5 environments
Source/License	PolyHeaven / CC0	Provenance
Tags	wood, dark, floor	Up to 20 per material

Synthetic Spectroscopy: Class Definition Example

$X_{test}$2

MatSyn25: Synthesis Record (Excerpt)

$X_{test}$3

---

References:

• (Li et al., 1 Oct 2025) Li et al., "Material Synthesis 2025 (MatSyn25) Dataset for 2D Materials"
• (Schuetzke et al., 2022) Schuetzke et al., "A universal synthetic dataset for machine learning on spectroscopic data"
• (Vecchio et al., 2024) Vecchio et al., "MatSynth: A Modern PBR Materials Dataset"