SemiConLens: Visual Analytics for 2D Semiconductor Discovery

SemiConLens is a visual analytics system that integrates a novel correlation-aware multivariate imputation method (CAMI) with interactive visualizations to overcome challenges related to data scarcity and reliability, thereby enabling materials researchers to effectively discover and evaluate promising candidates for 2D semiconductors.

The Gist

Imagine you are a treasure hunter searching for a very specific type of gold in a vast, chaotic ocean. In the real world, this "gold" is a new type of 2D semiconductor—a material only a few atoms thick that could power our future computers, phones, and energy grids.

The problem is that the ocean is huge, the water is murky, and the map you have is full of holes.

This work introduces SemiConLens, a high-tech diving suit and sonar system designed to help scientists find these materials faster and more reliably. Here is how it works, broken down into simple concepts:

1. The Problem: A Map with Missing Pieces

Scientists know that conventional computer chips are hitting a limit; they are becoming too small and overheating. They need new 2D materials to fix this.

• The Old Way: Scientists used to mix chemicals in the lab, test them, and hope for the best. This is like trying to find a needle in a haystack by poking the hay one by one with a stick. It is slow and expensive.
• The New Way (The Challenge): They began using computers and Artificial Intelligence (AI) to predict which materials would work. But there is a catch: the AI is like a student who hasn't read enough textbooks. The data is "sparse" (full of missing pages), and the AI often guesses without knowing if it is right. This is like asking a weather forecaster to predict next week's rain when they only have data for three days.

2. The Solution: SemiConLens

The authors developed a system called SemiConLens, which serves as a bridge between the raw power of AI and the human intuition of the scientist. It has two main engines:

Engine A: The "Intelligent Filler" (CAMI)

Since the data map has holes, the system must fill them without inventing false facts.

• How it works: Imagine you are trying to guess the price of a house in a neighborhood where some price tags are missing. Instead of guessing randomly, you look at the neighboring houses that are similar (same size, same style) and see what they sold for.
• The Innovation: The system uses a method called CAMI. It examines how different material properties relate to each other (correlation) and finds similar materials (similarity) to fill in the missing numbers. It is like a detective using contextual clues to solve a puzzle rather than just guessing.

Engine B: The "Crystal Ball with a Trust Meter" (Prediction and Uncertainty)

Once the map is filled, the system uses AI to predict how good a material will be. But unlike a standard AI that outputs only a single number, SemiConLens tells you how confident it is.

• The Analogy: If a normal AI says, "This material is 90% efficient," SemiConLens says, "This material is 90% efficient, but I am only 60% confident in this number because I had to guess some of the data."
• Why this matters: This prevents scientists from wasting time on materials where the AI is merely "hallucinating" (guessing wildly).

3. The Interface: A Dashboard for Discovery

The system presents this data on a screen with three main views that feel like a cockpit for treasure hunting:

• The Filter View (The Sieve): Here, scientists set their rules. They can say, "Show me only materials that are stable at room temperature and have high speed." This is like adjusting a sieve to catch only the gold nuggets and let the sand fall through. It also shows a "trail of history" so they can see how they arrived at their current list.
• The Discovery View (The Star Map): This is the coolest part. The materials are displayed as circular badges (glyphs).
  • The inner pie slice: Shows the key statistics (such as speed and energy).
  • The outer ring: Shows why the material has these statistics (which tiny atoms cause the effect).
  • The Layout: The badges are arranged so that similar materials group together like stars in a constellation. The system uses a special algorithm to push them apart so they do not overlap, even if there are thousands of them.
• The Comparison View (The Scorecard): When a scientist finds two promising candidates, they can place them side by side in a heatmap. This is like comparing two cars on a datasheet, but with a 3D model you can rotate to see the molecular structure.

4. Does it work?

The authors tested this system with real materials scientists (experts in this field).

• The Verdict: The experts loved it. They said it helped them find potential materials much faster than before.
• Real-World Test: In one test, they used the system to find materials for splitting water into hydrogen fuel. The system quickly identified a material (WS2) that experts already knew was good, proving the system works. In another test, they found candidates for a new type of energy-saving transistor.
• A Minor Complaint: The system is somewhat slow (taking about 7 seconds to update) when thousands of elements are on the screen, as it performs heavy mathematics to prevent the badges from overlapping. But the experts said the trade-off was worth it.

Summary

SemiConLens is a tool that takes a messy, incomplete dataset of potential materials, uses intelligent mathematics to fill the gaps, and then displays the results in a way that allows human experts to recognize patterns, verify the AI's confidence, and make better decisions. It transforms the chaotic ocean of data into a clear, navigable map for discovering the next generation of technology.

Technical Summary

Technical Summary: SemiConLens: Visual Analytics for the Discovery of 2D Semiconductors

Problem Statement
The discovery of new two-dimensional (2D) semiconductor materials is crucial for overcoming the performance limits of traditional silicon-based transistors at the nanometer scale. However, this process faces significant hurdles. Traditional experimental workflows are slow, costly, and heavily reliant on trial-and-error. Computational approaches, such as Density Functional Theory (DFT), are accurate but computationally intensive and slow for large-scale screening. Machine Learning (ML) offers a faster alternative but is hindered by two main problems in materials science: (1) Data Scarcity, where datasets like C2DB contain thousands of compounds but only a few hundred with labeled values for critical attributes (e.g., CBM GW, VBM GW); and (2) Reliability and Trustworthiness, where ML models often act as "Black Boxes," providing no transparency regarding prediction uncertainty or the reasons behind decisions.

Methodology
The authors propose SemiConLens, a visual analytics system designed to bridge the gap between human domain expertise and ML capabilities. The system operates through two integrated modules: a scarcity-aware prediction pipeline and an interactive visualization interface.

1. Scarcity-Aware Prediction Pipeline:

   • Correlation-Aware Multivariate Imputation (CAMI): To address data scarcity, the authors developed CAMI. Unlike standard imputation, CAMI fills missing values iteratively by first filtering out columns/rows with excessive missingness, followed by calculating a Pearson correlation matrix. For a target column with missing values, the most correlated columns are identified; their missing values are imputed using column means, and then a similarity-based row selection (using K-Nearest Neighbors) is employed to estimate the target value based on the most similar observed instances. This preserves inter-attribute correlations better than global mean imputation.
   • Model Selection and Uncertainty Estimation: The system evaluates multiple ML models (RF, SVR, MLP, GB, NGBoost, Autoencoder) to predict five key semiconductor attributes. Based on performance, Autoencoders (AE) are selected for electron and hole mobility (leveraging few-shot learning capabilities), while Natural Gradient Boosting (NGBoost) is used for GW bandgap, VBM GW, and CBM GW.
   • Uncertainty Quantification: Uncertainty is explicitly calculated: for AE via reconstruction error (mean absolute percentage error); for NGBoost via relative standard deviation (RSD), derived from predicted standard deviations. Feature importance is determined using SHAP values (SHapley Additive exPlanations).

2. Visual Analytics System:
   The interface consists of three coordinated views:

   • Filter View: Provides an overview of attribute distributions using parallel coordinates and embedded scatter plots to visualize uncertainty. It includes a control panel for attribute selection and a tree diagram for tracking the exploration history.
   • Discovery View: Visualizes the compound space using a novel circular glyph design. The inner pie chart encodes user-selected key attributes, while the outer circular bar chart encodes influencing factors (feature importance or correlation strength) to explain why a compound exhibits certain values. To manage visual clutter in dense clusters, a cluster-aware layout optimization is employed. This combines cartogram scaling (enlarging areas based on cluster size) with a force-directed algorithm that includes boundary containment and position retention forces to minimize overlaps while preserving cluster topology.
   • Comparison View: Enables detailed analysis via an interactive heatmap where rows represent compounds and columns represent attributes. Data sources (predicted, imputed, ground truth) and uncertainty (via circle radius) are visually encoded. It also integrates the visualization of 3D molecular structures.

Main Contributions

• CAMI Method: A novel imputation technique that utilizes both attribute correlation and sample similarity to handle sparse material datasets, outperforming mean and zero imputation in prediction tasks.
• SemiConLens System: An end-to-end visual analytics framework integrating scarcity-aware prediction with interactive visualization. It uniquely combines ML predictions with uncertainty quantification and explainability (via SHAP) to support trustworthy discovery.
• Visual Design Innovations: The introduction of a two-layered circular glyph that simultaneously displays attribute values and their influencing factors, paired with a cluster-aware layout optimization algorithm that preserves cluster structures while reducing glyph overlaps in high-density visualizations.

Results and Evaluation
The system was evaluated through quantitative metrics, expert interviews, and two use cases.

• Quantitative Evaluation: The CAMI method significantly improved prediction performance for electron and hole mobility compared to mean and zero imputation. Model evaluation confirmed that AE models were superior for predicting mobility, while NGBoost performed best for bandgap-related attributes.
• Expert Interviews: Twelve domain experts (PhD candidates, postdocs, and research fellows) evaluated the system. Participants rated the system very positively regarding effectiveness (mean 6.00–6.67) and usability (mean 6.50). Experts highlighted the system's ability to identify promising candidates, as well as the usefulness of the filter history and uncertainty visualization. Some feedback indicated that information density could be overwhelming and that system response times (approximately 7 seconds due to layout optimization) could be improved.
• Use Cases:
  • Water Splitting: The system successfully identified WS2 as a candidate similar to the benchmark MoS2 for hydrogen evolution, validating the workflow of filtering and cluster exploration.
  • Tunnel Field-Effect Transistors (TFETs): The system supported the identification of donor-acceptor pairs with specific type-II band alignment and demonstrated the ability to combine computational filtering with domain-specific expert knowledge (e.g., excluding non-layered structures).

Significance
The work claims that SemiConLens effectively assists materials researchers in discovering high-potential 2D semiconductors by making the discovery process more transparent and reliable. By explicitly addressing data scarcity via CAMI and visualizing prediction uncertainty and feature importance, the system enables informed decision-making in an area where data is often limited and traditional methods are too slow. The work demonstrates that combining human intuition with ML-driven insights, mediated through a tailored visual analytics interface, can accelerate the identification of desired materials for next-generation technologies.