1. Introduction
In the vast expanse of the universe, our planet has been a remarkable cradle of life for billions of years. It has navigated through countless challenges since its inception, particularly in the last three hundred years of the industrial revolution, when humanity finally broke free from the long-lasting poverty trap. In the meantime, we have also incurred escalating costs, including income inequality, climate change, biodiversity loss, and the depletion of limited nonrenewable resources. Conventional scientific methodologies and approaches are increasingly coming to their limits as we try to better understand the high complexities of the human-planet relationship, and to steer through the planetary health challenges.
2. PHAS Framework
As a response, in October 2023, Peking University launched a groundbreaking project to construct a Planetary Health Axis System, or PHAS, an AI-driven large dynamic global digital system, aiming to track human civilization footprints in relation to the well identified Planetary Boundaries. In order to accomplish such a huge challenging task, the PHAS builds on a four major axes system, including human health, species health, environment health and societal health, which are quantified by over 48000 key variables across the planet. In the following, this video presents a brief introduction of the PHAS project.
3. PHAS Roadmap
PHAS builds a roadmap using AI to model the comprehensive health of the planet. The PHAS roadmap consists of four core modules as shown.
Module 1 presents a stochastic control optimization process for the ultimate objective function of human civilization with trade-offs among the multi-targets of development. It brings a paradigm shift from focusing on GDP growth to human sustainable development for infinite time horizon. Instead of discounting future welfares, PHAS treats each generation equally by adding them into an infinite time average summation, following ergodic theory to ensure its convergence.
Module 2 builds a large data training platform to integrate all available data and current knowledge using iterative AI imputation models equipped with special trained LLM. In order to avoid the “hallucination problem” in LLM, PHAS incorporates a vertically fine-tuned pre-trained large model to extract prior information from literature, and to synergistically use data and human expertise through iterative training cycles.
Module 3 develops the PHAS core engine using deep causal neural network on graph, empowering the capacity to model the complexity of planetary health. It is a combination with non-invertible dimensionality expansion (nIDE) and scalable deep causal graph learning(SCGL). The nIDE helps to transform nonlinear complexity into high-dimensional linear complexity asymmetrically to ensure causality. The SCGL detects the connection network among these complexities.
Module 4 constructs an expert-data-AI convergence system to cease effective overfitting regularization of complex algorithms with real experts. PHAS generates unit reaction pathway results from statistics on current trained model and using decomposition tools to facilitate systematic quantitative and qualitative expert verification. Critically, the knowledge generated during these processes is absorbed in the network, leading to an ongoing learning system.
4. PHAS Outperform
Concerning the performance of PHAS model, it indeed offers a great deal of advantages over popular models in predicting capabilities. For instance, an evaluation of point prediction and graph recognition was conducted across both synthetic and authentic data-sets, demonstrating greater performance over conventional approaches such as statistical models, generic general large language models, general complex networks, localized natural science models, and general equilibrium models. In the context of loss function, the performance is measured by the accuracy on the predicted value with respect to the real or set value.
As indicated in the left panel, following the Y axis measured by a normalized accuracy scale for experimental value, experimental connections and real world value respectively, the results well demonstrate the superiority of the PHAS Full Model by the higher bars over most of the compared models. In terms of average scales on the X axis in the right panel, the overall performance of PHAS model outperforms the other methods.
In special, we test the model performance of predicting the distribution of climate change, especially on temperature. Different moments and quantiles are tested to measure the performance. The result shows that given similar but a bit better on the mean prediction, PHAS performs much better on high order moments and tail quantiles.
Moreover, performs on time decay loss are tested in a special case of food water use. Comparations are taken among the accuracy of predictions on different time length, say 1 year to 20 years. Results show that with the highest average accuracy, PHAS Full model provides lowest relatively time decay loss. This means the PHAS model performance would not decay much as looking deep into the future.
5. Weighting System & PHAS Matrix
Based on its multidisciplinary science foundation, PHAS provides a comprehensive planetary health metrics using a well-grounded weighting system.
Unlike traditional data-driven health index measures, PHAS derives its Planetary Health Metrics, for both the globe at the top layer and regionals at lower layers, from its system foundation of optimality. Based on the previously introduced modeling foundation, the weighting system of PHAS is derived from the optimal solution to a dynamic control problem aimed at maximizing humanity's ultimate objective function. Such a weighting system is further enhanced by iterations on well-defined database and human feedbacks from multidisciplinary panel of leading experts. The PHAS-derived Planetary Health Metrics presents human’s optimal expected ultimate objective function achievable from the current state at the current time point.
In general terms, the top-level measure provided by PHAS can be interpreted as a scalar value ranging from 0 to 100, where 100 represents the optimal level of the ultimate objective function expected to be achieved by humanity on a completely unpolluted, undamaged, and unconsumed planet, while 0 corresponds to the expected termination of human civilization. This measure offers a relative indication of the current overall health status of the planet and humanity.
The PHAS simulation functions permit scenario analyses and other policy evaluations. Using the Planetary Health Metrics as an example, where Y axis is normalized from 0 (the worst) to 100 (the best). X axis represents the yearly timeline, PHAS model indicates how the global level of planetary health would continue to deteriorate without actions today, shown in Panel 1.
6. Policy Lab
Alternatively, a 20% reduction of fossil fuel use would slowdown the current trend, but not a halt of it indicated in Panel 2. Similarly in Panel 3, a carbon emission tax increase from a world average of 0.3% to 5% would also help reduce the downward sloping of planetary health, but still not to alter the trajectory. However, if the global innovation investment can increase over 20%, it would lead to a reversal of the trend back to the pre-2000 levels as demonstrated in Panel 4.
7. Replicate & Go Beyond
In terms of other applications, PHAS is fully capable of replicating known scientific knowledge while yielding more systematic additions with stronger significance. Conventional approaches establish point-to-point relationships between elements through limited data analysis under specific conditions or assumptions. PHAS can generate consistent findings in a single integrated process.
Taking a case of climate change and human health, PHAS employs interpretability-oriented approximation to manifest conditional causal graphs for generating comprehensible pathway network. Virtually all known scientific discoveries can find corresponding representations in this network, such as temperature effects on disease prevalence.
Moreover, beyond conventional science, the PHAS model offer additional capabilities. For example, PHAS can identify longer-range and more complex correlation pathways, and its discoveries exhibit bidirectional exclusivity. This implies the PHAS-generated network constitutes a comprehensive mapping from origins to endpoints with inherent significance. Such advantages by PHAS enable comparative assessment and priority settings across the system pathways and key elements, providing relative pathway flux to help optimize policy making.
8. Call for Collaboration
PHAS is still on its second year, being a learning and growing AI system. Active utilization and systematically captured user-feedback will be an essential part of the interactive learning system and we do look forward to engaging with you to make PHAS an important and useful tool to help steer mankind through the overarching challenge of our time. Thank you.
