Machine Learning Based Identification of Solvents from Post-Desiccation Patterns

This paper presents an optimized protocol using artificial neural networks to identify solvents in starch-liquid slurries from post-desiccation fracture patterns with 96% accuracy by analyzing nine morphological features, particularly crack area distribution.

The Gist

Imagine you are a detective trying to solve a mystery, but instead of looking for fingerprints or footprints, you are looking at cracks in dried mud.

This paper is about a new, super-smart way to figure out what kind of liquid was used to make those cracks, even if the liquid has completely evaporated and vanished.

Here is the story of how they did it, broken down into simple steps:

1. The Setup: The "Drying Mud" Experiment

The scientists took a bowl and mixed cornstarch (like the stuff you use for gravy) with different liquids: water, ethanol (drinking alcohol), and acetone (nail polish remover). They also made mixtures, like 20% alcohol and 80% water.

They poured this "mud" into a flat dish and let it dry. As the liquid evaporated, the cornstarch shrank and cracked, forming beautiful, complex patterns on the surface.

The Mystery: If you look at the dried pattern today, can you tell if it was made with water or alcohol? To the human eye, it's tricky. The patterns look similar, just slightly different in size or shape.

2. The Detective Work: Measuring the Cracks

The researchers didn't just look at the pictures; they measured them like a forensic scientist. They used a computer to break every crack down into numbers. They looked at:

• Size: How big are the cracked pieces?
• Shape: Are they round, long, or weirdly shaped?
• Neighbors: How many other pieces is each piece touching?
• Order: Do they look like a neat honeycomb or a messy pile?

They turned all these measurements into histograms. Think of a histogram like a bar chart at a concert showing how many people are wearing red, blue, or green shirts. Here, the "shirts" are the sizes of the cracks.

3. The Brain: The Artificial Neural Network

This is where the magic happens. The scientists fed these "concert charts" (the histograms) into a computer brain called an Artificial Neural Network (ANN).

Imagine you are teaching a child to recognize animals. You show them a picture of a cat and say, "This is a cat." Then you show a dog and say, "This is a dog." Eventually, the child learns the patterns (ears, tail, fur) without you explaining the biology.

The computer did the same thing. It looked at thousands of crack patterns and learned:

• "Oh, when the cracks are small and the 'red shirt' bar is high, that's water."
• "When the cracks are huge and the 'blue shirt' bar is high, that's acetone."

4. The Big Discovery: The "Smoking Gun"

The team tested many different combinations of measurements to see which ones helped the computer guess the liquid best.

They found a "Golden Rule": The size of the cracks is the most important clue.

• Water makes tiny, intricate cracks (like a spiderweb).
• Alcohol and acetone make much larger, chunkier cracks.

When the computer focused heavily on the crack area, it became incredibly good at guessing.

5. The Result: A 96% Success Rate

By combining the crack size with a few other simple measurements, the computer achieved a 96% accuracy rate.

This means if you handed them a dried, cracked cornstarch plate and asked, "What liquid made this?" the computer would be right almost every single time, even if the liquid was a tricky mix of water and alcohol.

Why Does This Matter?

Think of this as a universal translator for nature.

• In Science: It helps scientists understand how different materials dry and crack.
• In Engineering: It could help figure out what went wrong with a dried coating or a dried paint job.
• In Space: The paper mentions that similar patterns exist on Mars and the Moon. If we find dried mud on Mars, this method could help us figure out what kind of ancient liquid (water? alcohol?) might have been there billions of years ago.

In a nutshell: The scientists taught a computer to "read" the scars left behind by drying liquids. By focusing on the size of the scars, the computer can now identify the invisible liquid that caused them with near-perfect accuracy.

Technical Summary

1. Problem Statement

The paper addresses the challenge of identifying the specific solvent used in a desiccation (drying) process after the liquid has completely evaporated. When colloidal suspensions (specifically cornstarch slurries) dry, they form complex fracture patterns. While the morphology of these patterns is known to depend on the solvent's properties (e.g., surface tension, volatility), visually distinguishing between solvents—especially in binary mixtures (e.g., water-ethanol)—is difficult and subjective. The authors aim to develop an optimized, automated protocol using Artificial Neural Networks (ANNs) to classify these patterns and identify the solvent with high accuracy, even when the solvent is no longer present.

2. Methodology

Experimental Setup

• Materials: Cornstarch-liquid slurries with an initial concentration of $\phi_0 = 1/2$ (equal mass of liquid and dry cornstarch).
• Solvents Tested:
  • Pure Solvents: Water, Ethanol, Acetone.
  • Binary Mixtures: Water-Ethanol mixtures at 20%, 40%, 60%, and 80% ethanol concentrations.
• Process: 40g of slurry was deposited in a 100mm diameter aluminum container. The system monitored mass loss, temperature, and humidity until complete evaporation (reaching ~20g of dry residue).
• Data Collection: 75 samples of pure solvents and 50 samples of binary mixtures were prepared. Digital images of the final fracture patterns were captured from a top-down view.

Morphological Analysis & Feature Extraction

The authors utilized image processing (FIJI/ImageJ and MATLAB) to extract nine morphological features from the fracture patterns, normalized by characteristic lengths derived from the container size and total crack length:

1. Size Features:
   • Normalized Pattern Area ($A_N$)
   • Normalized Perimeter ($P_N$)
   • Normalized Crack Area ($A_{crN}$)
2. Shape Features:
   • Isoperimetric Ratio ($\lambda = 4\pi A/P^2$)
   • Eccentricity ($\varepsilon$)
3. Geometric/Ordering Features:
   • Number of Neighbors ($N_B$) derived from Voronoi diagrams.
   • Orientational Order Parameters ($\psi_4, \psi_5, \psi_6$) calculated based on the angles between adjacent cell centroids.

Machine Learning Framework

• Input Data: Frequency histograms of the nine features were generated for each sample.
• Model Architecture: A Multi-Layer Perceptron (MLP) neural network.
  • Input Layer: Feature histograms.
  • Hidden Layers: 1 to 10 layers.
  • Output Layer: Softmax function classifying the solvent type (3 classes for pure solvents; 5 classes including mixtures).
• Training Strategy:
  • Algorithm: Scaled Conjugate Gradient (SCG) backpropagation.
  • Validation: 5-fold cross-validation.
  • Architecture Optimization: The study tested 31 variants of neuron counts (60–120) and compared three architectures: increasing, constant, and decreasing numbers of neurons in hidden layers.
  • Feature Selection: The model was trained on individual features and various combinations (up to 511 possible sets) to identify the optimal subset.

3. Key Contributions

1. Optimized Neural Network Architecture: The authors demonstrated that a decreasing MLP architecture (where the number of neurons reduces in successive hidden layers) yields higher classification accuracy compared to increasing or constant architectures.
2. Identification of Critical Features: The study isolated the normalized crack area ($A_{crN}$) as the single most predictive feature for solvent identification.
3. Feature Set Optimization: The paper proposes a specific protocol for selecting feature sets to maximize accuracy while minimizing processing time and avoiding overfitting.
4. Robust Classification Protocol: A complete workflow is established that successfully identifies solvents in both pure and binary mixtures, even when visual distinction is ambiguous.

4. Results

Single Solvent Classification (Water, Ethanol, Acetone)

• Accuracy: Achieved >95% accuracy.
• Best Performance: A two-feature set consisting of the orientational order parameter ($\psi_4$) and crack area ($A_{crN}$) achieved 100% accuracy.
• Observation: Using all nine features led to slight overfitting and longer processing times without significant accuracy gains.

Binary Mixture Classification (Including Water-Ethanol Mixtures)

• Accuracy: The overall average accuracy for all five classes (3 pure + 2 mixtures) was 96(±1)%.
• Best Performance:
  • The four-feature set ($\lambda, \psi_4, \psi_5, A_{crN}$) achieved the highest accuracy of 97.0(±1.2)%.
  • The two-feature set ($\psi_4, A_{crN}$) achieved 96.8(±1.4)%, proving that a smaller, optimized set is nearly as effective as a larger one.
• Worst Performance: The order parameter $\psi_6$ alone yielded the lowest accuracy (51–59%).
• Overfitting: Using all nine features resulted in a drop in accuracy to 93.6(±1.6)%, confirming the necessity of feature selection.

Key Finding on Feature Importance

The crack area distribution ($A_{crN}$) was identified as the dominant factor. The study found that more volatile solvents (acetone, ethanol) produced significantly thinner cracks and different area distributions compared to water, making this feature highly discriminative.

5. Significance

• Scientific Impact: This work provides a quantitative method to reverse-engineer the chemical environment (solvent type) of a dried material solely based on its physical fracture pattern. This is valuable for fields ranging from materials science to planetary geology (e.g., analyzing Martian soil patterns).
• Methodological Advancement: The paper moves beyond simple pattern recognition by demonstrating that feature selection is as critical as the neural network architecture itself. It proves that a carefully chosen subset of features can outperform a "kitchen sink" approach (using all available data).
• Practical Application: The proposed protocol offers a fast, automated, and highly accurate tool for quality control in industrial processes involving drying slurries, or for forensic analysis where solvent identification is required after evaporation.
• Generalizability: The authors suggest this protocol can be adapted to optimize pattern recognition in other scientific and engineering fields where morphological classification is required.