Category Archives: Peer Reviewed Article Review

instructions for the AI scientist

Leave a reply

Here’s an example of how AI science could work. If you ask machine learning to review a data set from a system and predict the future behavior of the system, it can often do a good job, but the mathematical approach it is using to do that is mostly opaque. It is also divorced from any sense of the system structure, its interaction with the physical universe, and the physical laws governing it. But you can ask the computer to constrain itself within those known physical laws, and then it may be able to provide you insights on the structure and physical processes inside the system. So at that point, you and/or the computer should be able to form scientific hypotheses and test them against the data. This example from Water Resources Research is about soil moisture.

Coupling Intuitive Physics Into Deep Learning for Soil Moisture Flow Processes Learning

Soil water flow processes in the unsaturated zone support ecosystems and regulate water, energy, and biogeochemical cycles. Recently, deep learning (DL) approaches have significantly advanced soil moisture (SM) prediction tasks yet still challenging to interpret. It’s difficult to peer into the internal reasoning procedures of algorithms, let alone associate them with specific physical processes. Thus, DL alone is unlikely to satisfy soil hydrological modeling needs and cannot advance process understanding. Here, we present DPL-S (deep process learning for SM dynamics) approach, which couples intuitive physics into deep learning architecture as structural guidance to facilitate comprehensive surrogate modeling of soil water flow. DPL-S discretizes the SM state evolution into multiple sub-process effects (e.g., gravity, matric potential) at the intuitive physics level and abstracts them into format-specific and learnable tensors. By cascading state-action matrices in a differentiable end-to-end framework and enforcing penalties for physical inconsistencies, DPL-S enables a profound understanding of physical functions and scenes of soil water flow. Comprehensive numerical experiments including layered soil conditions and tests with in situ observations, demonstrate that it achieves reliable SM profile reconstruction with predictive performance comparable to the state-of-the-art DL model on supervised items. The internal inference of DPL-S is fully transparent and the tensor representations achieve strict physical realism under limited water content supervision, thus enabling continuous predictions like physical models during the testing period. The model’s flexibility, generalization, noise resistance, and large-sample diverse data synergies are also evaluated. This work represents a solid step toward learning hydrophysical processes from large data sets.

the future of scientific and engineering modeling

Leave a reply

I’ve been using AI to assist me with coding (R, in my case) since shortly after ChatGPT came out. In engineering, we tend to run off-the-shelf models that were of course written in some kind of code. Sometimes these are open access but often they are proprietary. The brutality of market discipline pretty much requires specialized off-the-shelf solutions in industry because customers are not going to be willing to pay for custom coding. The proprietary ones are even often preferred for legal/liability reasons. Anyway, the future of modeling appears to be humans providing a detailed specification to an AI agent which then follows it to do the coding, debug the coding, run the model, process and present the results. The humans have to be able to detect whether the results are BS, of course, at this point in history. One can imagine using a different agent or a more specialized agent to assist the humans in the bullshit-detection stage, that agent getting more independent over time, and so in a cycle. I wonder if we will be using agents to set up, run, and post-process the specialized models, or if things will trend toward just letting the agents write more fundamental code over time. Or maybe the specification will be what future scientists, engineers, and business people focus their efforts on, with translating that into 0s and 1s being basically a commodity done on the fly by AIs. This makes sense to me – the most crystal clear function of AI so far, in my view, is making it easier and easier for humans to communicate with computers in more abstract language, logic, and mathematical symbols.

Anyway, this example used something called Roo Code which included a couple versions of Claude Code along with some other agents, to run a fishery-related model. There is a peer reviewed article, but I also like this blog post and this example of a specification given to the agents.

MEMOP, MOP, and SMOP

Leave a reply

These are some more decision-making frameworks I hadn’t heard of, at least by name. I am hopeful that AI can help techniques like this make the transition from gown to town.

Developing Robust Management Pathways for Nutrient Pollution in Watersheds Under Climate Uncertainty

Nutrient management represents an enduring effort toward sustainability. However, long-term management planning faces notable challenges, mainly due to substantial investments required under uncertainty of forthcoming climate. To address these challenges, this paper proposes and tests a Multi-climate-scenario (MCS) Multi-epoch Multi-objective Planning (MEMOP) framework (in combination MCS-MEMOP). This framework divides the long-term planning horizon into multiple epochs, allowing nutrient mitigation measures (e.g., fertilization management, filter strip) to be initiated at any epoch, each with its own water quality and investment constraints. To tackle climate uncertainty, it incorporates principles of Robust Decision-Making. MCS-MEMOP generates solution pathways outlining the timeline and progression of management measures, tested here for a case of a small, agriculture-dominated watershed. Considering a single climate scenario, the MEMOP method was compared with Multi-objective Planning (MOP) and Stepwise MOP (SMOP) for a 25-year nutrient management horizon, using the SWAT model to evaluate the test case water quality effects of solution pathways. Results show that MEMOP’s multi-epoch approach generates a larger and more diverse set of solutions than MOP and SMOP, offering greater flexibility to select optimal trade-offs among objectives. Additionally, MEMOP solutions exhibit superior cost-effectiveness compared to MOP and SMOP solutions. Applied separately to different climate scenarios, the MEMOP results show that changed climate conditions may significantly alter the Pareto front. In contrast, MCS-MEMOP yields robust solutions that can consistently satisfy 72%∼89% of epoch-specific constraints under new climate conditions in the test case, with a cost increase of 12% that reflects the price of addressing climate uncertainty in this case.

pedestrian level of service

Leave a reply

I had actually never heard of pedestrian level of service, but it appears to be just applying traffic flow modeling principles to pedestrian flow. It is intuitively appealing to me because you can just add pedestrians, cyclists, or whatever mode you want to a transportation model and specify which links in the network are open to which types of “traffic”. Theoretically, you could try to optimize the total flow of people from where they live to the places they need to get to, and not just maximize the flow of motor vehicles. Surely someone must have looked at this. A valid criticism, of course, is that these models can be short-term focused, even looking just at a single weekday peak hour and certainly missing longer-term dynamics like how infrastructure capacity and land use policy affect trip generation over time. Another criticism is that this engineering approach completely misses the quality of life aspects of urban design.

Pedestrian infrastructure assessment: Walkability vs. pedestrian level of service

This paper explores two of the most explored indicators for evaluating pedestrian infrastructure: walkability and Pedestrian Level of Service (PLOS). Walkability, typically used by urban planners, emphasises the qualitative aspects of the built environment, such as safety, comfort, accessibility and aesthetics, whereas PLOS, primarily employed by transport engineers, quantifies the operational performance of infrastructure based on pedestrian flow, density and speed. A systematic PRISMA-based literature review was conducted, covering 60 PLOS studies and 55 walkability studies were analysed in terms of definitions, contributory factors, data collection methods and modeling techniques. Despite sharing the goal of promoting pedestrian-friendly environments, these frameworks differ fundamentally in scope, purpose and methodology and are often applied independently. The findings indicate that walkability indicators vary in how factors are measured and allocated across dimensions. Moreover, walkability is treated as a “static” factor, both conceptually and methodologically. Relatively limited research examines how walkability changes over time (e.g., day vs. night) or varies across population groups. Conversely, PLOS generally excludes socio-spatial dimensions, a choice consistent with its original purpose rather than a methodological limitation. Some approaches attempt to incorporate subjective factors, but usually in ways resembling traditional walkability metrics. This study highlights the need for greater standardization in definitions and assessment frameworks, while also identifying challenges that complicate their practical applicability. Their complementary use can significantly enhance the design, evaluation and planning of pedestrian infrastructure, supporting more livable, sustainable and inclusive cities.

Some facts and figures on how birds die

Leave a reply

This article on a site called ZME Science has some facts and figures on how many birds are killed (in the U.S., I think) by wind turbines compared to other causes.

  • wind turbines: 140,000-680,000 birds per year; 0.3 – 0.4 birds per GW-hr of power production
  • power lines: 12-64 million birds per year [Daddy, how can the birds sit on the power lines without being electrocuted? Well, I guess the answer is that the ones you see on the power lines are the ones that didn’t get electrocuted.]
  • vehicle collisions: 89 million – 340 million birds per year
  • glass buildings: 1 billion birds per year [well, the article says “up to almost”. Rounding is fine with me, but they are oddly precise on some numbers and willing to round on others.]
  • cats: 1.3 billion – 4 billion birds per year [I happen to like cats just fine, but if you consider yourself a friend of animals in general it is really kind of immoral and hypocritical to have an unsupervised outdoor cat.]
  • fossil fuels: 5.2 birds per GW-hr of power produced “through habitat destruction, mercury poisoning, and acid rain” [presumably, the study referenced here followed some pretty different methods than the ones that looked at birds running into stuff and getting disemboweled by cat claws. But still, you can say this is an order of magnitude more than the wind turbines per unit of energy supplied, controlling for the fact that fossil fuels are supplying more energy in an absolute sense.]

There are some things you can do to make the wind turbines less bad for birds. But really, it’s the cats and the cars and just the general wanton destruction/displacement of nature by our civilization. Now I’ve depressed myself, as I often do.

Decision Making Under Deep Uncertainty

Leave a reply

I like the idea of applying decision theory to real-world problems. The ideal concept is that decision makers and stakeholders have a body of highly relevant objective information available to them at the time they are also applying judgment and values (and politics) to make those decisions. There is a huge academic endeavor, but the hard part in my work experience is pulling together relevant information in a timely manner. Today’s schedule and budget expectations just don’t allow complex analyses in many cases, so we default to relatively simplistic approaches to complex problems. Maybe AI can help apply these methods more quickly, alongside more traditional methods like benefit-cost analysis. I like the focus on robust decisions in this paper, and I like the collection of practical tools including some R packages.

A review of tools and resources to support Decision-Making Under Deep Uncertainty

Decision-making under Deep Uncertainty (DMDU) offers approaches to support robust, adaptive strategies for complex decision-making. However, practical uptake of DMDU remains limited, partly due to fragmented access to resources and a lack of an inventory of available tools. This study introduces a comprehensive catalogue of tools and resources. Through a structured survey and expert elicitation, we identify 28 resources and 16 tools that support DMDU research and practice and classify them using an established DMDU taxonomy. Our analysis reveals a focus on introductory guidance regarding theory and methods of DMDU application, with some bias toward water-related applications. Technical, method-specific resources on how to implement existing frameworks remain limited. Our results identify tools supporting all core DMDU components, though they highlight persistent scalability challenges. The resulting online catalogue provides a foundation for expanding the use of DMDU in practice and is intended as a living, community-driven platform.

more on how tipping points could unfold

Leave a reply

I’ve suggested that the climate tipping point might only be called in retrospect, and that the year we pick might be 2025. Not because it can be pin-pointed that precisely, but because if we decide in retrospect that the 2020s were about when it happened, that will be a nice round number to pick.

Here is one scenario from OneEarth journal on what a cascade of tipping points could look like.

The risk of a hothouse Earth trajectory

Warming from greenhouse gas emissions accelerates Arctic sea ice and Greenland Ice Sheet loss, reducing albedo and adding meltwater that weakens the Atlantic Meridional Overturning Circulation (AMOC). A weakened AMOC shifts tropical rainfall patterns, increasing drought risk and potential dieback in the northern Amazon forest, further amplifying global warming through the feedback involving carbon loss. Note that once one tipping point is crossed, it will likely impact the timing and temperature thresholds for other tipping points.

habitat area and fragmentation

Leave a reply

I gave a talk this week on a niche topic involving plant selection for stormwater management features like rain gardens in cities. I had just one slide on habitat connectivity and fragmentation as an interesting area for further research. That one slide generated a lot of interest. And it is an interesting topic. First of all, it has been looked at quite a bit in the design of nature reserves, but not so much in urban areas or areas with a mix of wild, agricultural, and urban land uses. And second, there is always the debate about whether focusing too much on connectivity metrics can detract from just preserving enough total habitat. The article below is an entry relevant to this question, relevant in the context of forests. In urban areas, in my view, this question gets flipped on its head. Fragmentation and disconnection is the de facto state, and can only be reversed on the margins. So the interesting question to me is what policy choices can make it the least bad. There is also the possibility that better policy choices in urban areas can reduce friction of (animal and plant) movement between wild landscapes, and even whether they can serve as relatively biologically functional islands in depleted agricultural landscapes.

Why controlling for habitat amount is critical for resolving the fragmentation debate

The need for a consensus on the effects of fragmentation per se is increasingly recognized (Miller-Rushing et al., 2019; Riva, Koper, et al., 2024; Valente et al., 2023) because deforestation continues and small forest patches are particularly vulnerable to destruction (Riva et al., 2022). If fragmentation per se reduces biodiversity, then policies should prioritize protection and restoration of large patches. If not, then policies should include all forest, irrespective of patch sizes (Riva & Fahrig, 2023). This would allow effective biodiversity conservation, even in human-dominated regions where no large patches remain, by protecting and restoring sufficient forest over a network of many small patches (Arroyo-Rodríguez et al., 2020).

what’s new with learning curves?

At least since reading some early Singularity-adjacent publications by Vernor Vinge, Ray Kurzweil, and Bill Joy, I’ve been interested in learning curves. (And for the record, the topic and these authors were were not considered politically “right wing” or even political at all at the time.) Progress, at least in certain technologies, tends to be exponential over time. This clearly applies to computer technology, where there are short product cycles, the needed infrastructure is more or less in place and/or can adapt as the technology is scaled/commercialized, and legal and institutional barriers to change are relatively low. Technologies with “recipes”, like chemicals, drugs, seeds and other agricultural technologies, might be other examples. For these we have the patent system to actually try to slow down scaling and commercialization to the pace of innovation. With energy technology, learning curves seem to play out over much longer periods of time because while available technology changes rapidly, our system tends to be locked into long-lived infrastructure that can only change slowly. So new energy technology rolls out slowly as it is scaled up and commercialized over decades. There are also entities with enormous political and propaganda power that fight tooth and nail to keep us locked into obsolete technologies and infrastructure that fit into their historical (and profitable) business models. Now when you get to other technologies, like transportation and housing, public policy, legal and institutional barriers are dominant and tend to retard or even prevent progress. Rollout is so hard that while there are pockets of innovation, many don’t see the light of day or don’t spread from the local/pilot scale, even if they are successful at this scale. These also vary by location and jurisdiction, so that progress is very uneven geographically. These are my thoughts anyway. As for what’s new, here’s a journal article from Advances in Applied Energy.

Variability of technology learning rates

Climate and energy policy analysts and researchers often forecast the cost of low-carbon energy technologies using Wright’s model of technological innovation. The learning rate, i.e., the percentage cost reduction per doubling of cumulative production, is assumed constant in this model. Here, we analyze the relationship between cost and scale of production for 87 technologies in the Performance Curve Database spanning multiple sectors. We find that stepwise changes in learning rates provide a better fit for 58 of these technologies and produce forecasts with equal or significantly lower errors compared to constant learning rates for 36 and 30 technologies, respectively. While costs generally decrease with increasing production, past learning rates are not good predictors of future learning rates. We show that these results affect technological change projections in the short and long term, focusing on three key mitigation technologies: solar photovoltaics, wind power, and lithium-ion batteries. We suggest that investment in early-stage technologies nearing cost-competitiveness, combined with techno-economic analysis and decision-making under uncertainty methods, can help mitigate the impact of uncertainty in projections of future technology cost.

This blog in Construction Physics has a deeper dive across more industries, and discusses at least one large data set that is available for analysis. If you could accurately predict learning rates (and successful scaling/commercialization rates) for specific technologies based on known factors, then theoretically you could fine-tune policies and incentives to increase the rate of progress in the technologies you want. So this is an area of research that could boost all other areas of research and progress.

forecasting extinction risk

I agree with this article that it doesn’t make sense to start protecting species only after they become rare and threatened. Forecasting which ones will become rare and threatened in the future could make sense. Of course, serious efforts to protect, create, and connect habitats would make the most sense. The method I am familiar with, which is appeals to me most, is the geographically-based metapopulation method of Ilka Hansky. But there are some others mentioned here that are new to me, or at least unfamiliar names for concepts I might have come across.

Forecasting extinction risk for future-proof conservation decisions

Conservation prioritisation emphasises currently threatened species, but there are strong arguments for complementary, more proactive approaches based on forecasting future extinction risk for unthreatened species. Forecasting methods vary in the timescale of extinction risk estimation and include established methods such as Population Viability Analysis (PVA) and Early Warning Systems, and emerging ‘Over-the-Horizon’ (OTH) methods. We develop a framework that integrates extinction risk assessment across timescales and outlines tradeoffs between shorter- and longer-term extinction prevention goals. This framework facilitates use of extinction risk forecasting in decision-theoretic conservation prioritisation that explicitly considers alternative time horizons for extinction prevention. Considering extinction risk on extended timescales offers a future-proof approach to conservation planning that may prevent more extinctions than focusing exclusively on currently threatened species.