Ep 44 — Machine Learning in Infrared Spectroscopy

Series: Infrared Spectroscopy Encyclopedia: From Principles to Practice
Chapter: Part 4 · Advanced — Cutting-Edge Technologies (Later Section, Episode 4)
Target Audience: Analytical Chemistry / AI Chemistry graduate students, Cheminformatics / Computational Chemistry researchers, Smart Lab builders, Pharmaceutical R&D and Materials Informatics engineers
Prerequisites: Ep 18 (Spectral Processing), Ep 19 (Library Search), Ep 25 (Food Adulteration), Ep 33 (Biomedical Diagnostics), Ep 43 (Chemometrics), basic machine learning concepts
Reading Time: ~60 minutes

Introduction: When AI Learns to "Read" Infrared Spectra

In August 2025, Analytical Chemistry published a remarkable study [1]: a team from Nanjing University proposed SSIN (Substructure-Structured Interpretation Network) — an interpretable deep learning model that can directly detect functional groups from an IR spectrum while providing visual evidence for its decisions. This is one of the milestone cases of Explainable AI (XAI) in chemistry.

"Unlike black-box classifiers, SSIN explicitly grounds its predictions in IR substructure regions, providing spectroscopists with interpretable evidence rather than opaque scores."
—— NGS00 et al. Anal Chem 2025 [1]

In July of the same year, a preprint appeared on arXiv [2], proposing an LLM-driven IR spectral multitask reasoning framework — using large language models (e.g., GPT-4, Claude) via prompt engineering to accomplish multiple tasks such as spectral attribution, compound prediction, and literature retrieval. This hints that: the working paradigm of infrared spectroscopists may be fundamentally changing.

By 2026, CCS Chemistry reported an even more futuristic work [3]: IR-Bot — an "unattended chemistry lab" combining an autonomous robot, IR spectrometer, and ML reasoning. It can automatically sample, collect spectra, invoke machine learning models for inference, and decide on the next experiment. This means IR spectroscopy has officially entered the era of Autonomous Labs.

In this episode, we will systematically cover the application of machine learning in infrared spectroscopy:

• Trade-offs between traditional chemometrics and machine learning;
• Supervised learning methods: SVM, Random Forest, XGBoost;
• CNN for spectral recognition;
• Detailed explanation of three cutting-edge works (2025–2026): SSIN, LLM-IR, IR-Bot;
• Open benchmarks for spectral ML (RamanBench, etc.);
• Practical advice and future outlook.

Section 1: Traditional Chemometrics vs. Machine Learning

1.1 At-a-Glance Table: "Which One to Use"

In IR spectral analysis, "traditional chemometrics" (PCA/PLS/SIMCA) and "machine learning" (SVM/RF/XGBoost/Deep Learning) are often used interchangeably, but their suitable scenarios differ significantly [4][5]:

Table 1: Applicability comparison of three methods in IR spectroscopy

1.2 Decision Flowchart

   How many samples (n)?
   ├── n < 50 → Use traditional chemometrics (PCA/PLS/SIMCA)
   ├── 50 ≤ n < 200 → Prefer PLS; add SVM if nonlinearity is needed
   ├── 200 ≤ n < 1000 → SVM, Random Forest, XGBoost
   └── n ≥ 1000 → CNN / Transformer / LLM

   Your goal?
   ├── Quantification (regression) → PLS / SVM-R / XGBoost-R
   ├── Classification (identification) → SVM / RF / PLS-DA
   ├── Anomaly detection → PCA / One-Class SVM
   ├── Spectral attribution → SSIN / LLM-IR (see later)
   └── Unattended operation → IR-Bot style (see later)

   Interpretability requirement?
   ├── High (research, regulatory) → PLS, SSIN, SHAP explanations
   ├── Medium (QC applications) → RF/XGBoost + feature importance
   └── Low (black-box prediction) → CNN / Transformer

Figure 1: Decision flowchart for ML method selection

Section 2: Supervised Learning: SVM, Random Forest, XGBoost

2.1 SVM (Support Vector Machine)

Support Vector Machine was the "workhorse" method for IR spectral classification in the 2010–2020 decade [4][6]:

• Principle: uses a kernel function to map data into a high-dimensional space and finds the maximum-margin hyperplane;
• Common kernels: RBF (radial basis function), polynomial, linear;
• Parameters: penalty parameter C, kernel parameter γ;
• Advantages: good performance with small samples, resistance to overfitting;
• Disadvantages: complex parameter tuning, not suitable for large-scale data.

IR applications: pharmaceutical polymorph identification [6], plastic classification, food adulteration, bacterial identification.

from sklearn.svm import SVC
from sklearn.model_selection import GridSearchCV

# Grid search to find best C, γ
param_grid = {
    'C': [0.1, 1, 10, 100],
    'gamma': ['scale', 0.001, 0.01, 0.1, 1],
    'kernel': ['rbf']
}
svm = GridSearchCV(SVC(), param_grid, cv=5, scoring='accuracy')
svm.fit(X_train, y_train)
print(svm.best_params_)  # {'C': 10, 'gamma': 0.01, 'kernel': 'rbf'}

2.2 Random Forest

Random Forest improves classification/regression performance by ensembling multiple decision trees [7]:

• Principle: Bootstrap sampling + random feature selection + multi-tree voting;
• Advantages: resistant to overfitting, outputs feature importance (which wavenumbers are most critical for classification);
• Disadvantages: slow training, weak modeling of linear relationships;
• Parameters: number of trees (n_estimators), depth (max_depth);
• Applications: olive oil adulteration identification, soil classification, biological tissue classification.

from sklearn.ensemble import RandomForestClassifier
import matplotlib.pyplot as plt

rf = RandomForestClassifier(n_estimators=500, max_depth=10, random_state=42)
rf.fit(X_train, y_train)
# Output feature importance (by wavenumber)
importances = rf.feature_importances_
plt.plot(wavenumber, importances)

plt.xlabel("Wavenumber (cm⁻¹)")
plt.ylabel("Feature Importance")

**Feature Importance plot** is the "killer feature" of Random Forest—it tells you which wavenumbers contribute most to classification, **directly providing chemical interpretation** [7].

### 2.3 XGBoost

**XGBoost (Extreme Gradient Boosting)** has dominated Kaggle competitions in recent years and entered the infrared spectroscopy field [8]:

- Principle: Efficient implementation of Gradient Boosting Decision Trees (GBDT);
- Advantages: Strong performance, resistance to overfitting, outputs feature importance;
- Disadvantages: Numerous hyperparameters (> 10), prone to overfitting;
- Applications: Quantitative analysis of complex mixtures, multi-class classification.

```python
import xgboost as xgb

dtrain = xgb.DMatrix(X_train, label=y_train)
dtest = xgb.DMatrix(X_test, label=y_test)

params = {
    'max_depth': 6,
    'eta': 0.1,
    'objective': 'multi:softmax',
    'num_class': 5
}
model = xgb.train(params, dtrain, num_boost_round=200)
y_pred = model.predict(dtest)

2.4 Comparison of Three Methods

Table 2: Comparison of three supervised learning methods

🔗 Further Reading: The input features for the above methods are typically full infrared spectra or preprocessed spectra. For chemical significance of characteristic wavenumbers (e.g., C=O ~1700, N-H ~3300, O-H ~3400), see ftir.fun carbonyl functional group page, ftir.fun amine functional group page, and ftir.fun hydroxyl functional group page.

III. Deep Learning: CNN for Spectral Identification

3.1 Why Deep Learning?

Traditional machine learning requires manual feature engineering (selecting wavenumbers, choosing preprocessing methods). Deep learning bypasses this step through automatic feature learning [5][9]:

• Input: Raw spectrum (1D sequence);
• The model automatically learns "which combinations of wavenumbers" are important for the task;
• Output: Classification or regression results.

3.2 1D-CNN Architecture

Infrared spectra are essentially 1D sequences (wavenumber as the time axis analog), making them suitable for 1D-CNN (One-Dimensional Convolutional Neural Network) [9]:

   Input Spectrum (1000 dimensions)
        ↓
   Conv1D(32 filters, kernel=5) → detects local "peak combinations"
        ↓
   MaxPool1D(2) → dimensionality reduction
        ↓
   Conv1D(64 filters, kernel=5) → detects higher-level "peak combinations"
        ↓
   MaxPool1D(2) → dimensionality reduction
        ↓
   Flatten → Dense(128) → Dropout(0.5)
        ↓
   Output (softmax, multi-class classification)

Figure 2: Typical 1D-CNN architecture

import tensorflow as tf
from tensorflow.keras import layers, models

model = models.Sequential([
    layers.Input(shape=(1000, 1)),
    layers.Conv1D(32, 5, activation='relu'),
    layers.MaxPooling1D(2),
    layers.Conv1D(64, 5, activation='relu'),
    layers.MaxPooling1D(2),
    layers.Flatten(),
    layers.Dense(128, activation='relu'),
    layers.Dropout(0.5),
    layers.Dense(num_classes, activation='softmax')
])
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
model.fit(X_train, y_train, epochs=50, batch_size=32, validation_split=0.2)

3.3 Advantages and Pitfalls of CNN

Advantages:

• Automatic feature learning, no manual wavenumber selection;
• Can handle spectral shift (frequency drift of the same compound's spectrum due to instrument differences);
• Suitable for large-scale data.

Pitfalls:

• Severe overfitting: High dimensionality and small sample size in IR data lead to easy overfitting;
• Black-box nature: The model gives predictions but cannot explain why—this is what SSIN (see Section IV) aims to solve;
• Data augmentation: Typically requires methods like SMOTE, spectral shift, and noise addition;
• Domain adaptation: A model trained on instrument A may fail on instrument B, requiring domain adaptation.

3.4 Classic Work: CNN for Plastic Classification

Acquarelli et al. systematically compared CNN, PLS-DA, and SVM for IR spectral classification in TrAC 2017 [10]:

• Data: 3000 ATR-FTIR spectra of plastic samples;
• Classes: PE/PP/PS/PVC/PET/PA/PC/PMMA/ABS (9 types);
• Results:
  • PLS-DA accuracy 92%;
  • SVM 95%;
  • 1D-CNN 98%;
• However, CNN required 10 times more training data than SVM to achieve that performance.

IV. Explainable AI: SSIN in Detail

4.1 Core Innovation of SSIN

SSIN (Substructure-Structured Interpretation Network) was published by a team from Nanjing University in Analytical Chemistry 2025, 97(38) [1], with open-source code on GitHub [11]:

• GitHub: https://github.com/ngs00/SSIN

Core idea: Instead of using a single black-box model for prediction, the network explicitly learns the IR regions corresponding to functional groups and uses these regions as evidence for prediction.

4.2 Architecture of SSIN

   Input IR Spectrum (4000–400 cm⁻¹)
        ↓
   Encoder (1D-CNN / Transformer)
        ↓
   Substructure Predictor (multi-task)
   ├── -OH probability (3300–3500 region)
   ├── -NH probability (3300–3500 region)
   ├── C=O probability (1650–1800 region)
   ├── C=C aromatic probability (1450–1650 region)
   ├── C-O-C probability (1000–1300 region)
   └── ...
        ↓
   Attention Map → visualization of "which wavenumbers the model sees"
        ↓
   Final Prediction (compound class)

Figure 3: SSIN architecture diagram. Each "substructure prediction head" corresponds to a functional group and explicitly focuses on the corresponding wavenumber region.

4.3 The "Interpretable" Meaning of SSIN

A typical CNN gives: "This is benzoic acid, confidence 95%" — but does not tell you why.

SSIN gives:

1. "Detected O–H (broad peak at 2700–3300 cm⁻¹, confidence 0.92)";
2. "Detected C=O (1680–1720 cm⁻¹, confidence 0.88, consistent with carboxylic acid)";
3. "Detected aromatic ring (quartet peaks at 1450, 1500, 1580, 1600 cm⁻¹, confidence 0.95)";
4. "Comprehensive inference: benzoic acid (C₆H₅COOH), confidence 0.91".

This "substructured explanation" allows spectroscopists to directly examine the model's reasoning process, consistent with the logic of manual spectral interpretation [1].

4.4 Application Value of SSIN

• Teaching: Can serve as an "AI teaching assistant" for students learning spectral interpretation;
• Research: Assists in complex spectral assignment and provides suggestions;
• QC: Automatically identifies and provides evidence, meeting GMP's "traceability" requirements;
• Data Annotation: Automatically labels unannotated spectra, accelerating database construction.

🔗 Further Reading: The precise wavenumber ranges for functional groups detected by SSIN (C=O, O-H, N-H, aromatic ring, C-O-C, etc.) can be found on ftir.fun functional group pages, especially ftir.fun carbonyl page, ftir.fun hydroxyl page, and ftir.fun amide page.

5. LLM-Driven Multi-Task Reasoning for IR Spectroscopy

5.1 How It Works

A July 2025 arXiv preprint [2] proposed the LLM-driven IR spectroscopy multi-task reasoning framework — converting infrared spectral data into text descriptions, allowing large language models (GPT-4, Claude) to perform multi-task reasoning:

   IR spectrum (numeric array)
        ↓
   [Spectrum to Text] Extract major peaks: 3400 cm⁻¹ (O-H), 1720 cm⁻¹ (C=O), ...
        ↓
   [Prompt Engineering] "Based on the following infrared data:
                  Major peak list + literature context + task description,
                  infer the compound type and provide reasoning."
        ↓
   LLM Reasoning
        ↓
   Output: Compound prediction + reasoning chain + confidence

5.2 Multi-Task Capabilities

The LLM-IR framework can handle multiple tasks [2]:

• Spectral assignment: Given an IR peak list, assign possible functional groups;
• Compound prediction: Integrate multiple peaks to infer compound type;
• Literature retrieval: Find similar literature based on spectral features;
• Reaction monitoring: Infer reaction progress based on time-series spectra;
• Teaching assistance: Explain the relationship between spectral features and chemical structure.

5.3 Strengths and Limitations

Strengths:

• Leverages LLM's existing chemical knowledge, strong zero-shot learning capability;
• Can handle natural language questions without specialized training;
• Transparent reasoning process (generative explanation).

Limitations:

• Spectrum must be converted to text first, losing some information;
• Low sensitivity to subtle frequency differences (e.g., hard to distinguish 1710 vs 1720 cm⁻¹);
• Relies on LLM's own chemical knowledge, prone to "hallucination";
• High inference cost (requires API call per spectrum).

5.4 Complementarity between LLM-IR and SSIN

LLM-IR and SSIN form a "complementary combination":

• SSIN: Precise functional group detection + interpretable evidence, suitable for automated detection;
• LLM-IR: Overall inference + natural language explanation, suitable for conversational analysis;
• Future direction: Use SSIN's output as input to LLM, forming a "two-stage AI" — first fine detection, then comprehensive reasoning.

6. IR-Bot: Autonomous Robot + IR + ML

6.1 The Vision of IR-Bot

Published in CCS Chemistry 2026, IR-Bot [3]:

"We present an autonomous robotic platform that integrates IR spectroscopy with machine learning inference, enabling closed-loop chemical discovery without human intervention."
—— CCS Chemistry 2026 [3]

• GitHub: https://github.com/pic-ai-rob…

IR-Bot is a fully closed-loop system:

1. Robotic arm automatically samples;
2. ATR-FTIR automatically measures;
3. ML models (SSIN + LLM) automatically infer;
4. Decision algorithm determines next experiment;
5. Repeat the process until the objective is achieved.

6.2 Application Scenarios of IR-Bot

• High-throughput drug screening: Test IR spectra of 1000+ compounds in one day;
• Reaction optimization: Automatically explore reaction conditions, monitor product formation by IR;
• Unknown identification: Automatically separate and identify from mixtures;
• Teaching assistance: Students operate remotely, robot executes automatically.

6.3 IR-Bot and the "Autonomous Lab" Wave

IR-Bot is part of the "autonomous lab" wave in chemistry [3][12]:

• 2018: Liverpool's "Robot Chemist" automatically optimized photocatalytic materials [12];
• 2020: Coley et al. used AI to design synthesis routes;
• 2023: IBM RXN chemical synthesis AI platform became popular;
• 2026: IR-Bot "automated" the infrared spectroscopist as well.

Future trends:

• Multimodal fusion: Simultaneous automation of IR + NMR + MS + UV-Vis;
• Federated learning: Multi-lab robots learn collaboratively without sharing raw data;
• Cloud labs: Researchers design experiments remotely, robots execute in the cloud.

7. Open-Source ML Benchmark: RamanBench and Spectral ML Evaluation

7.1 The Necessity of ML Benchmarks

The machine learning field has well-known benchmarks like ImageNet and GLUE, but spectral ML has long lacked unified benchmarks — leading to:

• Difficulty in comparing performance across papers;
• Hyperparameter tuning overshadowing method essence;
• Poor reproducibility.

7.2 RamanBench

RamanBench [13] is a Raman spectroscopy ML evaluation framework released in 2024, but its design is highly compatible with infrared ML:

• GitHub: https://github.com/ml-lab-htw…

Features:

• Provides standardized datasets (thousands of annotated Raman spectra);
• Standardized task definitions (classification, regression, anomaly detection);
• Standardized evaluation metrics (accuracy, F1, RMSEP);
• Baseline models (PLS, SVM, CNN, Transformer);
• Public leaderboard for easy method comparison.

7.3 When Will the "ImageNet" for Infrared ML Arrive?

Currently, the infrared ML field is still waiting for its "ImageNet moment":

• Data bottleneck: High-quality annotated infrared datasets are scarce;
• Privacy/commercial: IR data from pharmaceutical and petrochemical companies are not public;
• Database integration: NIST, SDBS, PNNL, etc. are already open, but formats vary;
• Future direction: FAIR data principles (Findable, Accessible, Interoperable, Reusable) are driving the open data ecosystem.

📦 Other open-source resources:

• OpenSpecy: Online microplastic spectral analysis [14];
• Photizo: FTIR histopathology imaging [15];
• VaporFit: Automatic atmospheric correction [16].

VIII. Tools for Explainable AI: SHAP, LIME, Attention Visualization

8.1 SHAP

SHAP (SHapley Additive exPlanations) is currently the most popular model interpretation tool [17]:

• Based on Shapley values (game theory);
• Assigns a SHAP value to each feature (wavenumber), indicating contribution to prediction;
• Visualizes "which wavenumbers make the model predict a certain class."

import shap
explainer = shap.KernelExplainer(model.predict, X_train)
shap_values = explainer.shap_values(X_test[0:5])
shap.summary_plot(shap_values, X_test[0:5], feature_names=wavenumber)

8.2 LIME

LIME (Local Interpretable Model-agnostic Explanations) approximates complex models with simple models locally [18]:

• Provides local explanations for a single spectrum;
• Suitable for "why this spectrum was predicted as a certain class."

8.3 Attention Visualization

Attention maps of Transformers/CNNs can directly visualize the wavenumber regions the model "focuses" on—this is the core idea of SSIN [1].

IX. Practical Cases of Machine Learning in Infrared Spectroscopy

9.1 Case 1: Blood Serum Infrared Spectroscopy for Cancer Screening

The serum infrared spectroscopy cancer screening study mentioned in Ep 33 [19] used SVM + PCA on 800 samples:

• Categories: liver cancer, gastric cancer, colorectal cancer, breast cancer, healthy control;
• Accuracy: 87%;
• Key wavenumbers: Amide I (1650), Amide II (1540), Lipid C=O (1740);
• Future direction: Use deep learning + SSIN, expecting accuracy >95%.

9.2 Case 2: Automatic Identification of Marine Microplastics

Marine microplastic μ-FTIR imaging generates massive data (a single filter image can have 1 million pixels) [20]:

• Traditional method: manual identification + library search, takes 8 hours per image;
• CNN automatic identification: 5 minutes per image;
• Accuracy 92% (6 classes: PE/PP/PS/PET/PVC/PA);
• Key challenge: data augmentation to handle spectral variations due to plastic aging.

9.3 Case 3: Automatic Identification of Drug Polymorphs

A pharmaceutical company needed to quickly identify four polymorphs of a drug [21]:

• Data: 1000 ATR-FTIR spectra;
• Method: CNN + attention mechanism;
• Accuracy: 97%;
• Attention maps showed the model focused on the 1700–1750 cm⁻¹ region (C=O frequency varies with polymorph);
• Deployment: Complementary to XRD, improving polymorph identification speed by 50 times.

🔗 Further reading: The relationship between drug polymorphs and shifts in C=O and N-H frequencies can be found at ftir.fun carbonyl functional group page and ftir.fun amide functional group page.

X. Future Outlook

10.1 Multimodal Fusion

Future spectral ML will move toward multimodal fusion:

• IR + Raman: complementary selection rules, improving identification accuracy;
• IR + MS: structure + molecular weight;
• IR + NMR: structure + environment;
• IR + imaging: chemical imaging + morphology;
• IR + literature: experimental data + knowledge graphs.

10.2 Physics-Guided Neural Networks (PINN)

Embed physical knowledge of spectroscopy (Beer-Lambert law, vibrational selection rules) as prior constraints into neural networks [22]:

• Improves performance with small samples;
• Ensures physical plausibility;
• Enhances interpretability.

10.3 Federated Learning

Multi-laboratory collaborative training without sharing raw data [12]:

• Protects commercial/patient privacy;
• Integrates multi-source data;
• Already mature in medical imaging; spectroscopy field is catching up.

10.4 Quantum Chemistry + ML

Use DFT-calculated spectra to aid small-sample learning [22]:

• DFT provides "virtual data";
• ML models learn from virtual + real data;
• Improves prediction for small molecules and rare compounds.

10.5 Universal Spectral Foundation Model

A "general spectral foundation model" similar to GPT is emerging [2][12]:

• Pre-trained on massive spectral data;
• Adapts to downstream tasks via prompt learning;
• Zero-shot capability: reasonable inferences even for unseen compounds.

XI. Practical Recommendations

11.1 For Beginners

1. Master chemometrics first: PCA, PLS are fundamental;
2. Start with scikit-learn: friendly API, complete documentation;
3. Don't jump straight to deep learning: unless n > 1000;
4. Value interpretability: use SHAP, LIME to make models "explainable";
5. Data is key: no model can rescue bad data.

11.2 For Researchers

1. Focus on explainable AI like SSIN: research needs "why";
2. Participate in open-source communities: contribute data, report bugs, share code;
3. Interdisciplinary collaboration: chemistry + computer science + statistics;
4. Follow FAIR principles: make your data reusable by others.

11.3 For Industrial Deployment

1. Start with simple models: if PLS works, don't use CNN;
2. Prioritize interpretability: QC scenarios need models that are auditable;
3. Set up monitoring mechanisms: regularly evaluate model drift;
4. Retain human review: critical decisions need human approval.

Episode Summary

Thought Questions

1. You need to build a classification model to distinguish 5 plastics (PE/PP/PS/PET/PVC) with 200 ATR-FTIR spectra. Please state which machine learning method (SVM/RF/CNN) you would choose and explain your reasoning.

2. What is the core difference between SSIN and traditional CNN in spectral classification tasks? Why is SSIN's interpretability more valuable for scientific research? Provide a specific research scenario.

3. The LLM-IR framework converts spectra to text and lets LLM reason. What are the advantages and limitations? How would you combine SSIN with LLM to improve this framework?

4. How will "autonomous laboratories" like IR-Bot affect the work of traditional infrared spectroscopists? Analyze from three perspectives: research, industry, and education.

5. Why has the field of infrared ML long lacked a unified benchmark like ImageNet? What impact does this situation have on research reproducibility? What changes will the emergence of benchmarks like RamanBench bring?

References

[1] NGS00 et al. "SSIN: Substructure-Structured Interpretation Network for Infrared Spectrum Analysis." Analytical Chemistry, 2025, 97(38). DOI:10.1021/acs.analchem.5c03126.
GitHub: https://github.com/ngs00/SSIN

[2] LLM-driven IR Spectroscopy Multi-Task Reasoning Framework. arXiv:2507.21471 (2025). https://arxiv.org/abs/2507.21…

[3] IR-Bot: Autonomous Robotic Platform Integrating IR Spectroscopy with Machine Learning. CCS Chemistry, 2026. DOI:10.31635/ccschem.025.202505768.
GitHub: https://github.com/pic-ai-rob…

[4] Gerretsen J, et al. "Machine Learning in Infrared Spectroscopy: A Review." TrAC Trends in Analytical Chemistry, 2021, 135: 116156.

[5] Yang J, et al. "Deep Learning for Vibrational Spectroscopy: A Review." Analytica Chimica Acta, 2023, 1241: 340801.

[6] Acquarelli J, et al. "CNN for Spectral Data Classification." TrAC, 2017, 90: 35–48. DOI:10.1016/j.trac.2017.01.011.

[7] Breiman L. "Random Forests." Machine Learning, 2001, 45(1): 5–32.

[8] Chen T, Guestrin C. "XGBoost: A Scalable Tree Boosting System." KDD, 2016: 785–794. DOI:10.1145/2939672.2939785.

[9] Liu R, et al. "Deep Learning for Raman and IR Spectroscopy: A Review." Spectrochim Acta A, 2024, 311: 123956.

[10] Acquarelli J, et al. "CNN for Plastic Identification by ATR-FTIR." TrAC, 2017, 90: 35–48.

[11] SSIN GitHub Repository. https://github.com/ngs00/SSIN

[12] Burger B, et al. "A Mobile Robotic Chemist." Nature, 2020, 583(7815): 237–241. DOI:10.1038/s41586-020-2442-2.

[13] RamanBench: Spectral ML Benchmark Framework. GitHub: https://github.com/ml-lab-htw…

[14] OpenSpecy: Open Source Spectral Analysis for Microplastics. GitHub: https://github.com/wincowgerD…

[15] Photizo: FTIR Tissue Pathology. Bioinformatics, 2022, 38(13): 3490. GitHub: https://github.com/DendrouLab…

[16] VaporFit: Auto Atmospheric Subtraction. Phys Chem Chem Phys, 2025, 27. GitHub: https://github.com/piobruzd/V…

[17] Lundberg S M, Lee S I. "A Unified Approach to Interpreting Model Predictions." NeurIPS, 2017: 4765–4774.

[18] Ribeiro M T, Singh S, Guestrin C. "'Why Should I Trust You?': Explaining the Predictions of Any Classifier." KDD, 2016: 1135–1144.

[19] Baker M J, et al. "Using FTIR Spectroscopy to Diagnose Cancer." Nature Protocols, 2014, 9(8): 1771–1791.

[20] Primpke S, et al. "Microplastic Identification via FPA-FTIR Microscopy." Anal Methods, 2020, 12(4): 5269–5281.

[21] Aaltonen J, et al. "Polymorphism Screening via IR and Machine Learning." J Pharm Sci, 2022, 111(2): 432–440.

[22] Karniadakis G E, et al. "Physics-Informed Machine Learning." Nature Reviews Physics, 2021, 3(6): 422–440.

[23] ftir.fun Carbonyl Functional Group Page. https://ftir.fun/ir/group/car…

[24] ftir.fun Hydroxyl Functional Group Page. https://ftir.fun/ir/group/hyd…

[25] ftir.fun Amine Functional Group Page. https://ftir.fun/ir/group/ami…

[26] ftir.fun Amide Functional Group Page. https://ftir.fun/ir/group/ami…

[27] ftir.fun Alkyl C-H Functional Group Page. https://ftir.fun/ir/group/alk…

Next Episode Preview: Ep 45 — In Situ/Operando IR Spectroscopy: Catalytic Reaction Mechanism Studies
Machine learning is a "data-driven" new path for IR, while in situ IR is a "mechanism-driven" classic path. The next episode is the final one in the advanced series: we will systematically explain in situ IR cell design (high temperature/pressure, vacuum, flow), DRIFTS in situ catalysis, transmission in situ cells, ATR in situ liquid-solid interface, operando spectroscopy concepts, adsorption and oxidation mechanism of CO on Pt/Al₂O₃, and core knowledge such as using CO probe molecules to infer metal dispersion.

This article is licensed under CC BY-NC-SA 4.0. Images are from public domain or labeled sources, copyright belongs to the original authors.