🤖 machine learning

Graph-Structured Hyperdimensional Computing for Data-Efficient and Explainable Process-Structure-Property Prediction

This paper introduces PSP-HDC, a graph-structured hyperdimensional computing framework that addresses data sparsity and heterogeneity in multiphoton photoreduction by encoding process-structure-property dependencies into intrinsic, explainable representations, achieving superior accuracy and generalization compared to conventional models.

Original authors: Jingzhan Ge, Ajeeth Vellore, Ajinkya Palwe, Ahsan Khan, David Gorsich, Matthew P. Castanier, SeungYeon Kang, Farhad Imani

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jingzhan Ge, Ajeeth Vellore, Ajinkya Palwe, Ahsan Khan, David Gorsich, Matthew P. Castanier, SeungYeon Kang, Farhad Imani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Building a House with a Tiny Toolbox

Imagine you are trying to predict how strong a 3D-printed house will be based on how you built it (the Process) and what it looks like under a microscope (the Structure).

In the world of advanced manufacturing, this is like trying to bake a perfect cake but you only have a handful of recipes, and every time you bake one, the ingredients are slightly different (some are wetter, some are hotter). The data is sparse (very little of it), messy (different scales and noise), and complex (changing the oven temperature doesn't just change the cake; it changes how the flour and sugar interact in ways you can't easily predict).

Traditional computer models are like students who try to memorize every single recipe they've ever seen. If they haven't seen a specific combination of ingredients before, they guess wildly or make up rules that don't actually work. They often find "spurious correlations"—like thinking that because you baked a cake on a Tuesday, it will always be fluffy, even though the day of the week has nothing to do with it.

The Solution: PSP-HDC (The "Smart Blueprint" System)

The authors created a new system called PSP-HDC. Instead of letting the computer guess the rules, they gave it a fixed blueprint (a graph) that shows exactly how the process leads to the structure, and how the structure leads to the final property.

Think of it like a flowchart for a detective:

The Blueprint (The Graph): The system knows that "Laser Power" affects "Pore Size," and "Pore Size" affects "Electrical Resistance." It cannot skip steps or invent fake connections. It forces the computer to follow the logical path of cause-and-effect.
The Translator (The Encoder): The data coming in is messy. One number might be a temperature (100 degrees), another might be a chemical weight (0.05 grams). The system uses a special "translator" that converts these messy, different-sized numbers into a universal language (high-dimensional vectors) so they can be compared fairly.
The Memory Bank (Associative Memory): Instead of calculating a complex equation, the system builds a "memory" for each type of outcome (e.g., "Good Conductive" vs. "Bad Insulator"). When a new sample comes in, the system asks: "Does this new sample look more like the 'Good' memory or the 'Bad' memory?"

How It Works: The "Lego" Analogy

Imagine you are building a complex structure out of Lego bricks.

Old Way: You dump all your bricks into a pile and try to guess the final shape by looking at the pile. If the pile is small, you might guess wrong.
PSP-HDC Way: You have a specific instruction manual (the Graph).
- First, you group your bricks by color (Process parameters).
- Then, you snap them together according to the manual to build a "Structure" block.
- Finally, you snap the Structure block onto the Process block to see what the final "Property" looks like.
- Because you followed the manual, you can't accidentally build a tower that defies gravity. The system only allows connections that make physical sense.

Why It's Better: The "Detective's Notebook"

One of the biggest problems with AI is that it's a "black box"—you know the answer, but you don't know why.

PSP-HDC is different because it keeps a detective's notebook.

When it makes a decision (e.g., "This sample will be conductive"), it can look back at its memory and say: "I decided this because the Laser Power was high, which created Small Pores, which led to Low Resistance."
It can break this down further: "Within the 'Pores' group, it was specifically the number of pores that mattered most, not their size."
This is called Explainability. It doesn't just give an answer; it gives the evidence trail, exactly following the blueprint.

The Results: Winning with Less Data

The researchers tested this on a real-world experiment involving 3D printing with lasers. They had a very small dataset (only about 60 samples).

The Competition: They pitted their system against standard AI models (like Random Forests, Neural Networks, etc.).
The Outcome: The standard models struggled. They overfit (memorized the noise) and failed when tested on new, unseen conditions.
The Winner: PSP-HDC achieved 91% accuracy on random tests and 89.6% accuracy even when tested on completely new manufacturing settings (a "process fold" test).
The Takeaway: By forcing the AI to follow the logical "cause-and-effect" graph, it learned the real physics of the process rather than just memorizing patterns. It worked better with less data because it didn't have to guess how the pieces fit together; the blueprint told it.

Summary for the Practitioner

If you are a factory manager or engineer:

Don't rely on black-box AI when you have very little data.
Do use a system that respects the known physics of your process (Process $\to$ Structure $\to$ Property).
PSP-HDC acts as a lightweight decision support tool. You tell it your process settings, and it predicts the outcome while showing you exactly which settings drove that prediction, helping you decide which experiments to run next.

Note: The paper explicitly states this is a decision-support tool, not a replacement for physical validation. It relies on you providing a correct "graph" (blueprint) of how your process works.

Technical Summary: Graph-Structured Hyperdimensional Computing for Data-Efficient and Explainable PSP Prediction

Problem Statement
The paper addresses the challenge of Process–Structure–Property (PSP) prediction in advanced manufacturing, specifically within the context of multiphoton photoreduction for 3D microstructures. The core difficulty arises from the "small-data, heterogeneous-statistics" regime:

Data Scarcity and Cost: Generating fully characterized PSP data points is expensive, requiring controlled fabrication, microscopy, segmentation, and property testing. Consequently, datasets are often limited to dozens of samples.
Interaction Dominance: The mapping from process parameters to properties is high-dimensional and interaction-dominated. Small changes in processing conditions can drastically alter microstructure and electrical performance through strongly coupled pathways.
Limitations of Existing Methods:
- Mechanistic Pipelines: Rely on calibrated submodels (e.g., thermal or kinetic models) that are often unavailable during early development and prone to error compounding when chained.
- Data-Driven Models (Black-box): Conventional feature-vector models (e.g., SVM, Random Forests, Deep Learning) are statistically underdetermined in this regime. They tend to learn spurious, regime-specific correlations rather than transferable causal dependencies, leading to poor generalization to unseen process conditions and unstable post-hoc explanations.
- Graph Neural Networks (GNNs): While GNNs encode relational structure, learning edge weights or attention patterns from tens of samples is statistically underdetermined, often resulting in brittle, regime-specific couplings.

Methodology: PSP-HDC
The authors propose PSP-HDC, a graph-structured Hyperdimensional Computing (HDC) framework designed to operate under these constraints. The methodology treats the directed PSP graph (Process $\to$ Structure $\to$ Property) not as an annotation, but as a fixed internal structural prior that dictates representation and inference.

Graph-Structured Representation:
- The PSP domain is defined as a directed graph where nodes represent scalar parameters (process or structure) and edges represent allowed dependencies.
- Composition: The framework uses HDC operators—binding (Hadamard product) and bundling (vector addition)—to compose representations.
- Graph Alignment: Directed dependencies are encoded by binding group hypervectors along graph edges. This ensures interaction terms exist only where the graph permits them, enforcing the process-to-structure-to-property ordering and preventing shortcut couplings.
Trainable Scalar-to-Hypervector Encoder:
- A key innovation is a trainable encoder that maps heterogeneous experimental scalars (with varying scales and noise) to hypervectors.
- Unlike standard HDC which uses fixed random projections for all variables, PSP-HDC learns parameter-specific embeddings ( $e_j$ ) on a fixed high-dimensional random basis ( $B$ ).
- This allows the model to calibrate node representations to the specific statistics of each measurement pipeline before graph composition, preserving HDC's noise tolerance while adapting to experimental heterogeneity.
Inference via Associative Memory:
- Class prototypes are formed by bundling the hypervectors of training samples within each class.
- Prediction is performed by retrieving the class prototype with the highest cosine similarity to the query sample's hypervector.
- This provides an explicit decision rule based on similarity margins.
Intrinsic Explainability:
- Explanations are derived directly from the same prototype memories used for prediction, avoiding post-hoc reconstruction.
- Multi-level Attribution: The model decomposes decision evidence into:
  - Parameter-level: Contribution of individual scalars.
  - Group-level: Contribution of physics-motivated groups (e.g., elemental composition, pore morphology).
  - Within-group: Relative contribution of parameters within a specific group.
- Memory Alignment and Separation (MAS): A new metric is introduced to quantify learning dynamics. It measures how well component memories align with their correct class prototype and separate from competing prototypes during training.

Key Contributions

Internal Structural Prior: The introduction of a domain-defined directed PSP graph as a computational constraint rather than an external annotation, converting causality into a structural prior for small-data learning.
PSP-HDC Framework: A graph-structured HDC model that enforces PSP topology through graph-aligned binding/bundling and performs prediction via prototype retrieval.
Adaptive Encoder: A trainable scalar-to-hypervector encoder that interfaces heterogeneous, noisy experimental data with the HDC framework via parameter-specific embeddings.
Graph-Aligned Explainability: Intrinsic attribution mechanisms at parameter, group, and within-group levels that provide coupled-effect evidence suitable for process planning.
Learning Diagnostics: The proposal of Memory Alignment and Separation (MAS) to quantify prototype formation and connect memory geometry to retrieval margins.

Experimental Results
The method was evaluated on a dataset of 60 specimens fabricated using the 3D OHMIC platform (multiphoton photoreduction of AgNO $_3$ ). The task was binary classification of sheet resistance (low vs. high).

Random Split Performance: PSP-HDC achieved a mean accuracy of 0.910 ± 0.077 over 1000 random 80/20 splits, outperforming strong baselines (SVM, Logistic Regression, Random Forests, CNN, GPC) which clustered around 0.875–0.885.
Process-Fold Generalization: Under regime holdouts (training on some process conditions, testing on entirely unseen ones), PSP-HDC achieved 0.896 mean accuracy. It remained stable across distribution shifts (e.g., different AgNO $_3$ concentrations or laser powers), whereas many baselines degraded significantly (e.g., dropping to ~0.5–0.6 in specific holdouts).
Learning Dynamics: Analysis showed that training sharpens class separation not just by moving prototypes, but by improving the encoding of samples toward their correct prototypes. MAS metrics confirmed increased alignment and separation at both parameter and group levels.
Explainability: Attributions revealed that low-resistance outcomes are driven by joint chemical composition and morphology, while high-resistance outcomes are driven by pore-geometry dispersion and process conditions, reflecting the coupled nature of the physics.

Significance and Claims
The paper claims that PSP-HDC offers a robust solution for data-scarce manufacturing problems where interaction topology is known but data is insufficient to learn it from scratch. By shifting from flat feature vectors to constrained relational composition, the method avoids overfitting spurious correlations and provides dependency-aware evidence rather than unstable post-hoc narratives.

The authors emphasize that PSP-HDC is a lightweight decision-support model, not a substitute for mechanistic validation. Its reliability depends on a well-constructed PSP graph and physically meaningful descriptors. The approach is presented as particularly valuable for early-stage process development, experiment prioritization, and automation problems where data is scarce but dependency structures are known. The work demonstrates that graph-structured HDC can recover transferable PSP coupling from small, heterogeneous datasets, outperforming both traditional mechanistic chains and standard data-driven models in regimes characterized by strong coupling and limited samples.