Abstract:
Reaching net-zero by cutting greenhouse gas emission by 2050 is arguably one of humanities greatest challenge. The sheer speed and scale at which this needs to be achieved brings into question whether such lofty goal can be achieved when only broad plans have been outlined. A significant part of the net-zero 2050 plan outlined by the IEA require more mining for minerals (e.g. for batteries), more geothermal energy (electricity + heating/cooling of megacities) and geological storage of CO2, to decarbonize industrial heat. With real world examples, I will argue that pulling off this challenge requires building intelligent agents to address the speed and scale issue. The specs of these agents are that they should be able to reason in high-dimensional physical, chemical, and geological spaces about uncertainty, interwoven with data acquisition and engineering operations. Two cases are used to illustrate the need for these agents and how they can be employed in a real setting. The first case concerns closing the estimated $12 trillion gap in battery metals discoveries needed as outlined in the EIA goals. The second concerns the complexity of storing CO2 in saline aquifers and depleted reservoirs under conditions that prevent leakages or earthquakes.
Bio:
Jef Caers received both an MSc (’93) in mining engineering / geophysics and a PhD (’97) in mining engineering from the Katholieke Universiteit Leuven, Belgium. Currently, he is Professor of Earth and Planetary at Stanford University, California, USA. Jef Caers’ research interests are decision making under uncertainty in developing the critical mineral supply as well as geothermal energy required to transition to 100% renewable energy. Jef Caers is founder of Stanford Mineral-X, a community building effort to strengthen stewardship for a prosperous future for all, powered by Earth's minerals. Jef Caers has published in a diverse range of journals covering Mathematics, Statistics, Earth Sciences, Engineering and Computer Science. Jef Caers authored or co-authored five books in the area of data science & decision making under uncertainty for Earth resources He was awarded the Krumbein Medal of the International Organization of Mathematical Geosciences for his career achievement.
Summary:
Goal: Transition society to 100% renewable energy
Wind&Solar
Water: Hydro power
Earth heat: Geothermal power
Earth’s minerals: Focus of this talk
Use-cases:
E.g Vienna:
Use temperature gradient to sub-surface (stable warmer or cooler temp) to heat/cool or generate power
Heat exchange requires pushing fluids underground
Need accurate models of geology to ensure there are no earthquakes
Carbon Capture and storage under ground in supercritical form
Must ensure stored carbon doesn’t leak or cause earthquakes
Need accurate models of geology of storage site
Circular materials economy
Key elements: Lithium, Copper, Nickel, Cobalt, rare earth elements
Need these to make Electric Vehicles
Effectiveness of mineral exploration has been declining 10x over the past few decades (e.g. no major copper mine discovered in past decade)
Challenges
US not ready to secure critical mineral supply
Takes a decade to bring mine to production
Unclear if we can create a resilient, decarbonized and just supply chain
Plans:
No Miracles Needed: https://www.cambridge.org/core/books/no-miracles-needed/8D183E65462B8DC43397C19D7B6518E3
We have all the tech needed
Heat pumps
Green H2
Etc.
Tesla Master Plan: https://www.tesla.com/ns_videos/Tesla-Master-Plan-Part-3.pdf
No need for fossil fuels
Nuclear is a promising technology
Need to coordinate the energy transition, make big decisions
Need AI to help plan this
Approach:
Sequential planning under uncertainty
Trade off safety, effectiveness, money
E.g. subsurface storage
Where to place monitoring instruments and which instruments?
How do we choose sites to inject?
E.g. mineral deposits
Focus: high-grade deposits (much more profitable to mine)
3-4% vs .2% mineral content
<1km deep, few 100s m wide
Analysis:
Aerial survey to understand coarse rifting structure
Then drill and measure to explore underground structure
Steps:
Belief: Geology state: dimension of state is enormous
Action: drill, inject
Observation: sensors with uncertainty
Reward
Driven by Markov decision process
Solved via Monte Carlo Tree search for sequences of steps
Search space explodes exponentially
Subsurface models
Needed to capture impact of actions and compare to observations
Full simulations are very expensive (need hundreds of thousands of runs)
Need deep learning surrogates to speed this up
Challenge: Physics depends on complex boundary conditions
Need 8k runs to train surrogate
Use importance sampling to ensure that training set has more interesting inputs on which the model may have problems
Bayesian Mineral Exploration
Focus on hypotheses and their falsification
Normally the approach is to tentative evidence of minerals and confirm it
Estimate posterior probability distribution of subsurface features
Intelligent prospector
Inject 10k documents on a given region
Use this to constrain the Bayesian posterior
Use Bayesian uncertainty to determine drill sites
Adaptively choose drill sites to minimize uncertainty
Normal: dense grid of drill sites
Intelligent design of mineral supply chain
From exploration to recycling
Focus on Lithium (Australia is primary source, followed by South America)
Supply chain models based on system dynamics