Improving the predictions of black carbon (BC) optical properties at various aging stages using a machine-learning-based approach

Baseerat Romshoo, J. M. Patil, Tobias Michels, Thomas Müller, Marius Kloft, Mira L. Pöhlker

TL;DR

This paper introduces a machine learning (ML) approach to accurately and rapidly predict the optical properties of black carbon (BC) aerosols, accounting for their complex fractal shapes and aging processes. By training ML models on a comprehensive database, the authors achieve significant speedups over traditional methods while maintaining high accuracy, paving the way for more realistic climate modeling.

What Problem Does This Paper Address?

Accurately estimating the climate impact of black carbon (BC) aerosols hinges on precisely knowing their optical properties (how they absorb and scatter light). Traditionally, BC particles have been simplified as spheres for these calculations, but this often leads to significant discrepancies with real-world observations. Real BC particles are complex, fractal aggregates formed from many smaller "primary particles," and they undergo aging in the atmosphere, acquiring coatings of other substances. Simulating these realistic morphologies with traditional physics-based models (like Lorenz-Mie theory or T-matrix methods) is computationally very expensive, limiting their use in large-scale climate models. This paper addresses the need for a fast, accurate, and computationally efficient method to predict BC optical properties that accounts for their realistic fractal nature and atmospheric aging.

Methods

The core of this work is the application of machine learning (ML) to predict BC optical properties. The authors developed a comprehensive database of simulated BC fractal aggregates with varying physicochemical properties (size, morphology, coating) and their corresponding optical properties, calculated using the multi-sphere T-matrix (MSTM) method. Two ML algorithms were trained on this database: Kernel Ridge Regression (KRR) and Artificial Neural Networks (ANNs).

Key aspects of their methodology include:

  • Feature Engineering: Careful selection of input features (wavelength, fractal dimension, coating fraction, particle size parameters) to avoid redundancy and optimize model performance.
  • Data Preprocessing: Application of the Box-Cox transformation to input features to improve model performance.
  • Model Training and Evaluation: Rigorous evaluation using random, interpolation, and extrapolation splits of the data to assess model generalization capabilities. Mean Absolute Error (MAE) was used as the primary performance metric.
  • Focus on Realistic Morphologies: The ML models are trained to handle fractal aggregate structures, moving beyond simplified spherical assumptions.
  • Inclusion of Aging: The models explicitly account for the "aging" of BC particles through the inclusion of a "fraction of coating" parameter.

Datasets

  • Simulated BC Fractal Aggregate Database: A newly constructed database containing 6192 simulated BC fractal aggregates. This database includes 35 features categorized into physicochemical properties, optical properties, and constants, covering various aging stages. This dataset is made publicly available.
  • Laboratory-Generated Soot Data: Data from laboratory experiments (Romshoo et al., 2022) were used to validate the ML model's predictions against real-world measurements.

Key Insights and Ideas

  • ML as a Game-Changer for BC Optics: The study demonstrates that ML can bridge the gap between computationally intensive physics-based simulations and the need for fast, accurate optical property predictions in climate models.
  • Fractal Morphology Matters: Incorporating realistic fractal morphologies significantly improves the accuracy of BC optical property predictions compared to traditional Mie-based methods.
  • Aging is Crucial: Accounting for the atmospheric aging of BC particles (represented by coating fraction) is essential for accurate optical property estimation.
  • Speed and Accuracy Synergy: The ML approach achieves predictions in fractions of a second, a massive improvement over the hours or days required by traditional methods, without sacrificing accuracy.
  • Generalizability: The trained ML models show good performance across various aging stages and physicochemical properties within the training data range.

Key Results

  • High Accuracy: Mean Absolute Errors (MAEs) for optical efficiencies were as low as 0.002-0.004, and for the asymmetry parameter, they ranged from 0.003-0.004.
  • Drastic Speedup: Prediction times are less than 1 millisecond per sample, compared to potentially over 24 hours for MSTM simulations.
  • Improved Predictions: ML-based predictions of single-scattering albedo (ω) and mass absorption cross-section (MAC) showed better agreement with laboratory measurements than conventional Mie-based predictions.
  • Validation: The ML algorithm's predictions compared favorably with laboratory-generated soot data, confirming the benefit of realistic morphologies.

Limitations and Future Work

  • Generalization Capabilities: While the models perform well within the training data range, interpolation and extrapolation experiments showed that performance can deteriorate when certain ranges are entirely removed from the training data, indicating limited generalization beyond the trained domain.
  • Coating Model: The study used a "closed-cell model" for coating, which is a simplification. More sophisticated coating models exist but are computationally more demanding.
  • Wavelength Dependence: The models treat wavelength as a continuous variable but were primarily trained on specific visible spectrum wavelengths. Testing generalization to other wavelengths would be beneficial but requires more data.
  • Future Work:
    • Extend the ML approach to include other morphological models of BC and different coating positions.
  • Develop a universal ML algorithm for all atmospheric aerosols, potentially integrating spherical particles and fractal aggregates.
  • Conduct closure studies using more extensive laboratory and ambient datasets.
  • Investigate Bayesian ML methods (like Gaussian Process Regression) for uncertainty quantification in predictions.
  • Explore the impact of other chemical compositions (e.g., inorganics) on optical properties.

Summary updated

Abstract

It is necessary to accurately determine the optical properties of highly absorbing black carbon (BC) aerosols to estimate their climate impact. In the past, there has been hesitation about using realistic fractal morphologies when simulating BC optical properties due to the complexity involved in the simulations and the cost of the computations. In this work, we demonstrate that, by using a benchmark machine learning (ML) algorithm, it is possible to make fast and highly accurate predictions of the optical properties for BC fractal aggregates. The mean absolute errors (MAEs) for the optical efficiencies ranged between 0.002 and 0.004, whereas they ranged between 0.003 and 0.004 for the asymmetry parameter. Unlike the computationally intensive simulations of complex scattering models, the ML-based approach accurately predicts optical properties in a fraction of a second. Physiochemical properties of BC, such as total particle size (number of primary particles (Npp), outer volume equivalent radius (ro), mobility diameter (Dm), outer primary particle size (ao), fractal dimension (Df), wavelength (λ), and fraction of coating (fcoating), were used as input parameters for the developed ML algorithm. An extensive evaluation procedure was carried out in this study while training the ML algorithms. The ML-based algorithm compared well with observations from laboratory-generated soot, demonstrating how realistic morphologies of BC can improve their optical properties. Predictions of optical properties like single-scattering albedo (ω) and mass absorption cross-section (MAC) were improved compared to the conventional Mie-based predictions. The results indicate that it is possible to generate optical properties in the visible spectrum using BC fractal aggregates with any desired physicochemical properties within the range of the training dataset, such as size, morphology, or organic coating. Based on these findings, climate models can improve their radiative forcing estimates using such comprehensive parameterizations for the optical properties of BC based on their aging stages.
BibTeX
@inproceedings{romshoo2024improving,
  author = {Baseerat Romshoo and J. M. Patil and Tobias Michels and Thomas Müller and Marius Kloft and Mira L. Pöhlker},
  title = {{Improving the predictions of black carbon (BC) optical properties at various aging stages using a machine-learning-based approach}},
  booktitle = {Atmospheric},
  year = {2024},
  doi = {10.5194/acp-24-8821-2024},
}
Figure