## Abstract

Atmospheric new particle formation (NPF), which is observed in many stations globally, is an important source

of aerosol particles and cloud condensation nuclei, affecting the climate directly and indirectly (Yao et al., 2018, Kerminen et al., 2018, Su et al., 2021). To obtain the particle number size distribution (PNSD) data for NPF studies, expensive

instruments and well-trained operators are required. Deep learning techniques such as the generative adversarial networks (GANs) (Goodfellow et al., 2014) provide possible tools to generate the PNSD data, leading to obtaining the PNSD data cheaply and conveniently.

Generally, GAN consists of a generator and a discriminator. The generator tries to find a function that can map

samples in one distribution A (e.g., the normal distribution) to another distribution B (e.g., a distribution whose samples

are images). Then the discriminator will check whether the distribution B is the same as the actual distribution C. As a

result, the objective of a GAN is to solve the minimax problem for the generator and the discriminator, and theoretically, the best solution can be found at the Nash Equilibrium.

We constructed a GAN, and generated new PNSD data through the PNSD data collected in the SMEAR II station

(Hyytiälä, Finland). Though the generated PNSD data do not have physical meanings, the generated data can help analyze the statistical properties of NPF events.

of aerosol particles and cloud condensation nuclei, affecting the climate directly and indirectly (Yao et al., 2018, Kerminen et al., 2018, Su et al., 2021). To obtain the particle number size distribution (PNSD) data for NPF studies, expensive

instruments and well-trained operators are required. Deep learning techniques such as the generative adversarial networks (GANs) (Goodfellow et al., 2014) provide possible tools to generate the PNSD data, leading to obtaining the PNSD data cheaply and conveniently.

Generally, GAN consists of a generator and a discriminator. The generator tries to find a function that can map

samples in one distribution A (e.g., the normal distribution) to another distribution B (e.g., a distribution whose samples

are images). Then the discriminator will check whether the distribution B is the same as the actual distribution C. As a

result, the objective of a GAN is to solve the minimax problem for the generator and the discriminator, and theoretically, the best solution can be found at the Nash Equilibrium.

We constructed a GAN, and generated new PNSD data through the PNSD data collected in the SMEAR II station

(Hyytiälä, Finland). Though the generated PNSD data do not have physical meanings, the generated data can help analyze the statistical properties of NPF events.

Original language | English |
---|---|

Pages | 23 |

Publication status | Published - 29 Oct 2021 |

MoE publication type | Not Eligible |

Event | 6th Finnish National Colloquium of Geosciences 2021 - Oulu University, Oulu, Finland Duration: 28 Oct 2021 → 29 Oct 2021 https://www.oulu.fi/katk/node/211918 |

### Conference

Conference | 6th Finnish National Colloquium of Geosciences 2021 |
---|---|

Country/Territory | Finland |

City | Oulu |

Period | 28/10/2021 → 29/10/2021 |

Internet address |

## Fields of Science

- 1171 Geosciences