- The global food supply must grow to meet demand.
- CybeleTech develops tools, software, and services to provide solutions.
- Curie supercomputer is setting the stage for a tech overhaul of agriculture.
The global agricultural industry faces tough challenges.
In order to meet the needs of a growing population, worldwide agricultural production must undergo an estimated 50% increase by mid-century. In its present state, however, it is already responsible for 70% of water consumed each year on earth and up to 32% of greenhouse gas (GHG) emissions.
In order to reduce this environmental impact, the hunt is on for new plant varieties that can produce better yields with fewer inputs.
Since it was founded in 2014, the French company CybeleTech has used its expertise to develop tools, software, and services to provide solutions to the challenges facing agriculture.
"We want farmers to be able to produce more with less, to give plants exactly what they require," says Denis Wouters, CybeleTech's resident expert in high-performance computing and data analysis. "There is a great need to identify new plant varieties that are more efficient, and we think that numerical technologies are one of the best ways to overcome these challenges."
Finding new varieties is indeed a challenge. Major seed companies already invest large sums in field trials, constantly experimenting with new varieties under various conditions to find the best ones. The whole process, however, can take up to 10 years. CybeleTech's goal is to substitute these field trials with numerical simulations of plant growth in order to help seed companies achieve superior varieties in a way that significantly reduces the cost to the environment, saving them both time and money.
Plant growth modeling requires some extremely complex calculations. "It is very expensive in terms of CPU time," says Wouters. "We ended up using all of the Curie supercomputer resources allocated to us through the SHAPE programme."
That is because the simulations that Wouters and the CybeleTech team have been working on take into account the intricate web of processes impacting plant growth. Field trials are comparatively simple, relying on yield observations, for instance, and then repeating the trials under different conditions.
But this is not the whole story. "If we only have the one observation," Wouters explains, "it is difficult to unfold all the processes to get parameters responding to each of those processes." What CybeleTech wants to do is simulate plant growth with models that take into account the plant's interaction with the environment.
The idea was to carry out a lot of simulations so that the project team could identify what protocols would be best for estimating genetic parameters. The same protocol can then be used to calibrate the plant growth model to currently available plant varieties.
In order to define these experimental protocols, there are three questions that need to be answered: What observables (yield, biomass, or leaf surface) should be measured? In what quantities? And in what environment?
"Our goal is to disentangle the interaction between the genotype and the environment," says Wouters. "This is the main strategy we are trying to introduce to the agricultural sector: the possibility to select varieties based on values that are independent of environment."
For instance, a farmer wants to determine the genetic parameters of a plant by using observations in the field. They can try to estimate the genetic parameters representing variety and then look at the difference in these parameters between different varieties.
Type A might have a growth rate that is higher than type B, so A will be better adapted in some conditions than B. In that case, the farmer would prefer to select type A for that environment.
Thanks to the power of the Curie supercomputer, the CybeleTech team managed to achieve an estimation of the parameter that surpasses the accuracy of the protocol used by major seed companies. Instead of a hundred trials per variety, it is only necessary to perform 10.
This advance is in no small part thanks to the experts at PRACE who were able to assist throughout the project. "They helped us to improve scalability and optimize the computation code itself, improving its speed by at least a factor of two and saving us half the time," says Wouters. "They did not just do it on their own but helped us to do it ourselves, so next time we will be more confident with this technology. The focus was on performance and optimizing it where you can."
But these skills are not needed just yet. After running the simulations, several months of data analysis have produced an optimised protocol that will enable seed companies to carry out fewer trials with greater accuracy. The next step for CybeleTech is to actually implement the protocol. When this is done, the data it produces can be used to write the model.
"Soon we should be receiving data from one of the major seed companies. Then we will provide them with the calibration tools and the tool that allows them to select from the genotypic parameters," Wouters says. "So now the big issue is to convince seed companies to actually make the shift to this kind of protocol, which is a big thing for them."
CybeleTech's primary objective is a technological overhaul of the agricultural sector. Wouters compares it to what happened in the automotive industry, when the introduction of numerical simulations revolutionized crash testing and allowed for more fuel-efficient vehicles.
It is never easy to change entrenched habits of a whole industry, but at its heart, this project has a simple selling point: numerical technologies mean you can produce more for less.
This project was awarded 4 million core hours on Curie hosted by GENCI at CEA, France.
This article is reprinted from the PRACE 2016 digest on HPC in Industry.
If you're headed to SC16 in Salt Lake City, stop by the PRACE booth (#3000) and see how they can meet your advanced computing needs.