Hello everyone. Welcome to our virtual chat room here for further discussions about the application of AI/ML to mineral exploration. We encourage everyone to share your opinions and thoughts here, whether you registered for the workshop or not, whether you are a novice or an experienced ML practitioner. Below is a list of questions that we have prepared to stimulate discussions.
How would you evaluate the current status of the use of ML in non-seismic problems?
What are the problems that you think can be better solved by machine learning (than traditional non-ML methods)?
• What is the potential impact of solving this problem with ML?
• What is the “doability” of this application? How long would it take to create this solution?
What are the challenges when applying ML to real-world problems?
What can we, as a community, do to make the best use of ML?
Do we need to create a few benchmark models and data for people to train, compare and evaluate different ML methods?
What advice would you give to a first-year geophysics graduate student who wants to study machine learning?
Who (or, which companies) are the leaders in applying ML to geoscience problems?
Can you provide a few ML references, including books, papers, tutorials, online courses, that you have found to be useful?
Please feel free to post other questions that you want to discuss with others.
Thanks in advance for your participation in the discussions! We hope to develop a workshop product based on these discussions. We hope this workshop product will summarize the recent advances, the state of the art and the road ahead in the mineral exploration community as well as the general non-seismic community.
Jiajia, here are a few suggestions for questions. Please pick any you think are useful for the wrap-up session.
How do you evaluate the robustness of different ML methods/services when used outside the domain of the training data? How can you achieve defensible outcomes with real geological problems? The outcomes are only as good as the model assumptions in the training datasets.
Is ML another form of inversion where you don’t have to compute the forward model or its derivatives?
Training ML methods on theoretical datasets does not guarantee that they produce robust results with real world geological problems. How do you modify the training procedures to manage these differences.
How do you manage the expertise cross-over between ML specialists that don’t understand the limitations of the training data and domain experts that don’t fully understand the limitations of ML methods. It is important to avoid the age old problem of garbage in – garbage out. Is healthy scepticism on both sides a necessary qualification?
Using domain experts to train ML systems for interpretive processing on real data appears to be a very promising area for productivity improvements. Do we need to build in QC processes the check results where non-skilled technicians take over the routine processing?
Are we at the stage of qualifying the many ML /AI methods for suitability in a range of application areas with attribution of quality of results, training cost, skill requirements and relative computing cost.
Can we convert the neural-net training model into a format that we can comprehend and directly modify to override limitations in the model training?