- contact@amc-lahti.fi
AI for atmospheric science and computer vision
AI for Atmospheric Science
This work addresses a significant challenge in atmospheric science, specifically the computational constraints faced by large-scale models. With thousands of chemical compounds interacting and outputs influenced by numerous environmental parameters, current air quality and climate models often rely on simplified chemistry schemes to mitigate the high computational costs associated with running full chemistry models. However, this simplification can introduce substantial bias into the results. An example highlighting this issue is a study conducted by Boy et al., where they observed a relative discrepancy of approximately 50% in methane lifetime under a scenario of a 6K temperature increase. The advancement of Artificial Intelligence (AI) has demonstrated significant progress across natural sciences, leading us to believe that AI holds promise in enhancing our comprehension of atmospheric chemistry.
Modelling atmospheric chemistry is complex and computationally intense. Given the recent success of Deep neural networks in digital signal processing, we propose a Neural Network Emulator for fast chemical concentration modelling. We consider atmospheric chemistry as a time-dependent Ordinary Differential Equation. To extract the hidden correlations between initial states and future time evolution, we propose ChemNNE, an Attention based Neural Network Emulator (NNE) that can model the atmospheric chemistry as a neural ODE process. To efficiently capture temporal patterns in chemical concentration changes, we implement sinusoidal time embedding to represent periodic tendencies over time. Additionally, we leverage the Fourier neural operator to model the ODE process, enhancing computational efficiency and facilitating the learning of complex dynamical behavior. Details can be found here.
Large-scale atmospheric air pollution models are computationally intensive due to numerically solving stiff systems of differential equations that describe chemical kinetics. Replacing numerical solvers with efficient neural network-based solvers has the potential to significantly reduce computational costs. Our initial focus towards modelling stiff atmospheric chemical kinetics is on estimating reaction parameters with neural networks. Understanding atmospheric chemical reactions is essential for analysing and predicting the concentrations of target species. However, estimating reaction parameters, such as rate coefficients, is particularly challenging due to limited experimental data. Neural network-based approaches have proven more efficient than conventional numerical methods in modelling chemical kinetics, but stiff chemical kinetics, where reaction rate coefficients span several orders of magnitude, still introduce obstacles. Based on existing approaches for kinetic parameter estimation and modelling stiff chemical systems, our work aims to develop methods to accurately determine rate coefficients for these challenging reactions.
Air pollution modeling is crucial for understanding and mitigating the impacts of pollutants. However, traditional methods, based on complex differential equations, are computationally expensive, especially at large scales or when using spatiotemporal data. Artificial intelligence (AI) offers a promising solution. Neural networks reduce computational costs while improving prediction accuracy, replacing numerical solvers to model complex chemical dynamics, including stiff kinetics. Our research targets two main objectives: (1) predicting pollutant concentrations across various regions and (2) accurately estimating chemical reaction parameters despite limited experimental data. AI models, trained on both simulated and real data, have proven effective in addressing these challenges. By integrating AI into atmospheric sciences, we aim to develop faster and more precise tools to improve air quality management and reduce pollution.
AI for Computer Science
Using Neural Networks to learn Image Super-resolution, we can restore the missing details, like edges and textures. Details can be found here.
A smart traffic monitoring system benefits from the surveillance camera and automatic AI vision recognition
We propose that FunnyNet-W simultaneously process images, audio, and text to understand the humor from video, just as humans do! Details can be found here.
Neural style transfer can mimic the desirable style to your photo, so that you can create your own artistic photo with any style. Details can be found here.
From LiDAR, we usually get sparse and noisy point cloud which affects the 3D vision tasks. Using Neural Networks to achieve point cloud upsampling/denoising can help autonomous driving and industrial design. Details can be found here.
We present See360, which is a versatile and efficient AI framework for 360-degree panoramic view interpolation using latent space viewpoint estimation. Details can be found here.
Contact Us
contact@amc-lahti.fi
Location
LAHDEN YLIOPISTOKAMPUS Niemen kampus Niemenkatu 73 (B-osa) 15140 Lahti