One uncomfortable reality about the shift to a Net Zero economy is that it will also require the discovery and mining of a lot of raw materials for the new infrastructure. Engines will be replaced by electric motors, generators with solar and wind, pipes with wires and transformers, and gas tanks with batteries. Building these will require discovering millions of tons of new minerals, many of which are inconveniently located owing to geopolitical issues or nearby communities.
Earth AI has announced some preliminary success in finding this buried treasure in new, previously unexplored areas. The mining industry calls these greenfields, in contrast to brownfield exploration adjacent to existing mines. The company recently discovered a rich deposit of molybdenum in Australia with twice the concentration of the largest existing mines.
Perhaps more impressively, its new approach achieved a success ratio of one in eight for greenfield discovery compared to the industry average of one in 200. Their results are even slightly better than traditional exploration techniques in brownfield exploration, with an average of one in 20.
This is a big deal since each one of those surveying holes costs a lot to dig. Earth AI has also pioneered a more systematic approach to digging these holes using modular components that help keep the costs down for each hole. Cheaper and faster holes also speed the loop on the experimental process to build a better AI model.
Earth AI CEO and founder Roman Teslyuk says:
Large mining companies have utilized AI to improve mining profits and prolong the life of mining operations. Start-ups like Plotlogic, MineSense, GoldSpot and Stratum AI have been at the forefront of this innovation. These start-ups have been successful at near-mine brownfield exploration and have a swath of large mining clients to whom they sell their services. What’s different about Earth AI is that we never tried to compete in a crowded data-rich mine data re-analysis. Instead, we have focused on finding new greenfield mineral deposits away from existing mines. It is much more challenging, but the reward is much, much larger.
Training an AI geologist
Looking for greenfield mining opportunities is a much more challenging problem. Brownfields are a well-studied and data-rich environment, while greenfields are data-poor. The standard and preferred solution is to collect new data, which is very costly and time-consuming.
The Earth AI team hypothesized they could replace the need for more data with a better approach for interpreting and labeling existing data. They trained their system on 400 million geological cases known from exploration archives. But this was no simple task. Teslyuk explains:
It is hard to focus the deep learning model on the right geological aspects at such a scale. We taught our AI to learn geology. Our AI acts like a geoscientist, studying each case, distilling the knowledge, and developing predictions. But this happens on a much bigger scale, generating consistent and dependable predictions.
The actual process consists of three phases:
- Targeting: The models are trained on the 400 million geological cases across Australia to identify areas of mineralization and highlight locations with a high probability of finding a mineral system. Teams go into the field to sample and review the targets.
- Hypothesis: The geologists study the mineral system on the ground. At this stage, a sister technology helps the geologists better understand the geological setting and aid in forming hypotheses.
- Drilling: The team tests its hypothesis by drilling down to 600-meter depth and proving or disproving the presence of mineralization. Each drill hole provides invaluable knowledge of the mineral system that is then fed back into the system and used to form new hypotheses.
Data quality is key
Earth AI recruited a rare mix of geology and deep learning experts. More importantly, it also adopted a research culture. Iterative experimentation helped distill important key components for a geological deep learning system to improve predictions. The company has been conducting research for six years and spent more than six hundred days in the field testing the predictions and gaining feedback for system improvement. Teslyuk observes:
The key challenge is researching the ways to focus the model on the right kinds of knowledge that need to be learned. The difficulty is that there is no rule book. Progress is only achieved by generating ideas, running experiments, observing and testing the model performance, and adjusting the system to make things work. We learned that we need to have a constant field presence to obtain real-world feedback.
The quality of data from each geological case is essential, but quality monitoring is a huge issue when working on a continental scale using hundreds of millions of data points. So, Earth AI built a half-automated expert-driven data review system that dramatically speeds up data quality review. For example, domain-specific software focuses on finding and remembering data errors and inconsistencies and fixing them at scale. Teslyuk says:
AI companies usually do not attempt such tasks, but we see this as a key ingredient of our success.
A systematic approach to drilling
Another factor was to consider drilling as an essential part of the data collection process rather than outsource it to existing contractors, as is common industry practice. This helped to optimize more steps in the process and streamline field operations.
The firm re-designed drilling hardware to make it modular, self-sufficient, and environmentally friendlier. This included integrating a waste-treatment system to reduce ground disturbance and environmental impact. They also streamlined logistics to carry more stock that was better organized for transport and subsequent analysis. The result is relocation between sites can be done within days, compared to industry norms of weeks.
The industry is known for its slow pace. To scale and speed up discoveries, you need to optimize and scale the hardware. Legacy systems are simply incompatible with predictable high-performing operations. We had to spend a lot of effort redesigning the hardware to streamline it with the specifics of the drilling operation.
As enterprises rush to add AI to their services and products, it's tempting to get caught up in the buzzy new AI models and services rushing to market. A much better approach is to focus on lowering the cost of creating and interpreting experiments, as Earth AI has done. In the long run, this will likely lead to better results as people figure out where these new tools improve accuracy and where they hallucinate.
This requires a cultural reframing from trying to create things faster to learning how to learn faster. As Earth AI’s experience has suggested, it took a lot of work to figure out how to automate the process of interpreting and structuring existing data and then confirm or negate each hypothesis more cost-effectively. This kind of shift will be required to succeed in solving the various problems necessary to build out Net Zero infrastructure.