Initially, spatial databases containing geochemical and flood risk information belonging to BGS, the Environment Agency and the Scottish Environment Protection Agency (SEPA) will provide national scale inputs into the AI tool(s). These will be supported then by targeted fieldwork with subsequent laboratory experiments and sample analysis to determine the drivers influencing contaminant mobility and bioaccessibility, which will be used to train artificial intelligence modelling tools. Key parameters are likely to be soil and water physiochemical properties such as pH and organic carbon. These factors highly influence aspects for example such as i) contaminant ion exchange with soil and sediments, ii) the dissolution of oxides (e.g., iron), which binds with contaminants, and iii) the formation of metal sulfide complexes. These factors all impact a contaminant’s mobility and bioaccessibility. These soil physiochemical properties differ spatially, influenced by soil texture, geology and mineralogy, anthropogenic inputs, and flood frequency. Such heterogeneities in spatial features make modelling changes in contaminant mobility and bioaccessibility complex, which is where artificial intelligence comes in to address these highly unstructured data.
To help validate the AI modelling output, at one location, a section of each field site will be partitioned off and inundated using pumped river water for a known period (3-5 days), with specific flow and depth parameters. Following a period of inundation, the residual floodwater will be pumped offsite. After a period of rest/drying (approximately 1 month) the site will be flooded for second time. Soil and flood water samples will be collected at up to ten sampling points across each test site pre- and post-flooding on both occasions. All samples will be analysed as below.
Building on previous work, we expect that soil samples from field sites across the UK will be subjected to simulated flooding under different extreme weather scenarios in the laboratory to test the AI model predictions for chemical remobilisation. Plastic cells filled with the air-dried soil samples will be inundated with simulated river water under different biotic and abiotic conditions (e.g., different soil temperatures or varying soil to solution ratio to simulate faster/slower flooding), followed by periods of drying and subsequent re-inundation. Saturation and drying times will reflect available values from the literature. Physico-chemical parameters will be monitored in real time and soils and inundation fluids will be sampled at each time-step and subjected to sample analysis as described below to determine contaminant mobility and availability for uptake.
Each sample from the laboratory or field will be subject to a range of analytical tests to determine how different PHEs (e.g., heavy metals, radionuclides, faecal indicator organisms) may be affected by the wet/drying regimes applied. Solid phase distribution of contaminants will be determined using sequential extraction. It is expected that analytical tests such as the following will be used to address the objectives.
• Real-time physico-chemical analysis using a Hana Flow Cell (DO, TDS, pH, ORP, Temperature), Infra-red spectrometry, particle size, organic matter, C, N analysis – to determine the influence of these parameters on contaminant mobility and availability
• Total elemental concentrations using ICP-MS or XRF – to determine PHEs present
Additionally, using the ISO standard  the potential human availability of PHE will be determined. The bioaccessibility protocol and sequential extraction data will be used to also determine the availability to certain wildlife species (e.g., mammals). The results of these analyses will be utilised to test the AI modelling.
Data processing and modelling
Using the data collected, the student will predict the environmental and human health impacts of contaminants mobilised and redistributed during flood events. Risk will be assessed by comparison with numeric criteria (e.g., soil guideline values, bathing water criteria etc.) and input into risk assessment modelling software such as PC Cream 2008  and CLEA  for humans, ERICA  for wildlife to determine the potential impact of PHEs. This risk predictor also needs to be combined with information on the flood probability along with the understanding the drivers for remobilising the PHEs and their associated bioaccessibility (obtained from artificial intelligence), along with information on population dynamics and people’s habits to produce risk maps.
We will further develop and test data driven AI models using traditional regression analysis or machine learning techniques to relate the data for the different PHEs and the co-factors such as temperature, water volume etc.
The resulting outputs will increase our knowledge of the behaviour and fate of contaminants affected by flooding and enable us to provide advice to people growing food for example in areas prone to flooding.