Machine Learning for Fluid prediction (post-stack seismic data, Innovation Norway grant 2020/521538)


In May 2020, PSS-Geo with partners (CaMa Geo, AGGS AS, and Lime Petroleum Noway AS) was awarded an Innovation Norway grant to extend the current research project of Rune Inversion (Artificial Intelligence-based seismic inversion algorithm) to fluid prediction. The produced inverted volumes of P-wave velocity, Density, P-impedance, and their algebraical extension to Volume of Clay and Porosity, created a unique possibility to predict a type of fluid using ML algorithms. 


Today, PSS-Geo is the holder of the unique dataset of seismic and well logs that gives the company a tremendous advantage in ML algorithms development for geo exploration.: 

  1. Seismic merge that is suitable for inversion (seamless, phase-controlled, with true amplitudes preserved)

    • Inverted seismic post-stack cube of 62 000km2: P-wave velocity, Density, P-impedance, Volume of Clay and Porosity

  2. Accurate CPI database with a unique "fresh-water" logs estimation (300 wells)

  3. Competence in Rock Physics, Seismic Processing, and Data science

If you are interested in joining the project, do not hesitate to contact us

The project participants have flexible possibilities to receive ML code, product (currently a prediction over the entire Norwegian Sea Shelf), and a product over the chosen area).  

The basic principal of the algorithm can be investigated in YouTube session of Fluxgate Technologies NG GeoPython Week: (3hours) 

The test area data (16 000km2), wells and the Python codes can be free downloaded from


The first result: dark blue - HC, yellow - brine.


Fluid prediction from seismic data using different machine learning methods. V. Kalashnikova, E. Karaseva, R. Øverås, T.R. Sharafutdinov. 2021. 7th EAGE Conference Tumen' - Extended Abstract (in rus).

Field examples. Ærfugl and Marulk gas/condensate reservoirs in the Norwegian sea.


Post-stack seismic data was inverted for P-wave velocity and Density using the Rune Inversion algorithm. The volume of clay was computed using Gyllenhamar's formula (from P-wave and Density), and porosity was computed using a standard approach. The produced volumes were used as recognition datasets, while 7 random wells' logs were used for learning. The results of the Machine Learning for fluid prediction are shown in Section 1 and Section 2.

Section 1


Section 2