Robots are doing?Science
and it's a good thing

Robots are doing?Science and it's a good thing

One research lab at a time, robots are taking over science.

They are discovering new drugs, novel methods to grow nanotubes and expanding our library of new materials.

They run complex experiments and tirelessly repeat measurements until they can confidently report to the last decimal digits. 

They keep meticulous records. They do not make errors or drop test tubes.

They don’t get tired or waste time wondering what the meaning of life is. Neither do they threaten to unionize or take sick leave.

Some of them are even beginning to formulate their own hypotheses and carry out experiments that they designed — to solve problems that they chose.

Meet Adam

Adam is one of the first robot scientists in the world. Currently he’s at Cambridge, UK where he runs experiments on yeast.

Yeast or Saccharomyces cerevisiae is the fungus responsible for wine, bread and beer. Like dogs and their breeds, yeast comes in a number of ‘strains’ each of which is slightly different genetically. Like an ambitious grad student, Adam hopes to run experiments on as many yeast strains as he can to reach his target objective.

Adam is shaped like a mobile clinic with robot hands, centrifuges, freezers and incubators. He does not have a face or a voice, but if you observe the video below carefully, you’ll see that he has a character — written surreptitiously into his careful movements, in the pauses between his actions and in the questions that he asks himself.

He selects specific strains of yeast from his freezer and transfers them to a microtiter — a plate with a multitude of small test tubes or ‘wells’. Inside each of these tubes is a nutritious sugary liquid with some selected additives.

Once every 30 minutes this plate is entered into a reader where the light reflecting off each well is carefully measured. Adam can use this information to determine how fast the yeast grows in each well. This tells him whether a particular chemical added to the well is beneficial or detrimental to the fungus.

If he chooses, he can go into one of these wells, scoop up some cells and let them grow in a different well, which he monitors as well.

Adam runs over a thousand yeast experiments simultaneously, each lasting upto 5 days.

 In this yeast Universe, Adam is a benevolent omniscient God. 


Through these experiments Adam hopes to find out which yeast genes are responsible for certain yeast enzymes.

While a giant leap for the yeast-kind, this study also has important implications for us.

Small single cellular entities like yeast are far easier to study than complex organisms like humans. Elucidating cellular mechanisms and information pathways from genes to enzymes in yeast can shed light on similar mechanisms in other species. In fact, several proteins important to us were first discovered as their close relatives (homologes) in yeast.



Adam is not just a mindless machine running one blind experiment after another.

From his massive database, he can pick out an enzyme and guess genes that might be responsible for it. He then runs targeted experiments to confirm or refute this hypothesis.

No human being is involved in this process. The team that runs Adam does not know which enzyme he is looking at or what hypothesis has been formulated. They only see the results once Adam has checked his hunches.

The initial results from Adam identified the genetic markers for over fifteen enzymes. 

But more importantly, it announced the birth of a new paradigm, that of the autonomous scientist.

No alt text provided for this image

Robot scientist Eve, located at the University of Manchester. Eve studies new drug candidates for diseases. Eve is a close cousin of Adam. 

While currently a small tribe, so great is the potential of these autonomous scientists that many believe that we are on the brink of a major revolution in science — one where our knowledge doubles every few years.

If so, this could be the beginning of the end of the ‘Edisonian’ approach in scientific exploration.



The Edisonian Paradigm

In 1878, Thomas Alva Edison was on the verge of discovering the first commercial electric bulb.

The physics was well known:

  1. Passing current through any material (ex: copper wire, silk, gas) heats it up. 
  2. Hot materials glow. 

If designed cleverly, these ubiquitous dual observations could be put together to make electric light.

To generate enough light however, these materials needed to be heated to extremely high temperatures. This meant that most materials either reacted with the air around it or turned into liquid, either of which is an undesired consequence.

Electric bulb

The light of the bulb is generated by the extremely hot filament at the center. The choice of the filament material is the result of massive scientific and engineering effort lasting years. Photo by Yuichi Kageyama on Unsplash

Edison’s challenge was to find the right material for his bulb that did not self immolate. At the minimum, the candidate material should be willing to pass current through it: ie, be a good conductor of electricity. This rules out clay, wood and glass.

Further, it should satisfy a number of common sense requirements. It should at least last a few days — even longer if possible. And it should not be so expensive that only Edison could afford it.

At that time however, there was no materials handbook that listed all materials in the world along with their optical, electrical, thermal, chemical and mechanical properties.

So Edison set out to physically test one material after another.

It is estimated that he experimented with over 2000 materials including his own hair until he landed on carbon filaments made from strips of the exotic oriental bamboo.



The Labor of Science

The dirty secret of modern science is that it still functions the same way that Edison pioneered some 150 years ago.

Sure, we have definitely come a long way in that much of this process is driven now by our deep scientific knowledge and intuition. We know the underlying physics and chemistry behind the things that we use or need, which helps to vastly narrow the search space. A scientist today is unlikely to add his hair into molten metal and hope for vibranium.

Science from the outside follows a highly organized logical progression where each step is the result of prior knowledge and contemplation.

But on the inside, there is still chaos and arbitrary choices.

No alt text provided for this image

Closed loop workflow of data driven science. The potential chemical space available for exploration is impossibly vast. Theoretical models and computer simulations help us narrow this down to smaller subset of combinations — which is too big for manual exploration but might be accessible to autonomous systems. The robotic agent studies a smaller subset of this region, measures the results and uses this information to improve its own selection and planning criteria. Image taken from ‘Using artificial intelligence to accelerate materials development’ 

To make a lightweight steel for example, the knowledge of the underlying physics and chemistry might tell a metallurgist that his best option would be to add nickel, chromium or manganese to Iron.

But from here, there is no real way forward unless each of these combinations is physically synthesized and the resulting alloy is characterized in a lab for its physical and chemical properties. These experiments could be straight forward or complicated. Repetitions might be required until we can be sure of the results.

At the ground level science is a human endeavor that is anchored in the unbiased observation of genuine natural phenomena.

 It doesn’t matter if one thousand calculations or equations point to nickel as being the best metal to partner with iron , as long as one observation refutes it.

Discovering new drugs, battery elements, parts for automobiles, glasses, solvents, textiles, ceramics, electronic materials, alloys, liquids, gels etc continue to be a laborious task requiring daunting manual labor.

This is where autonomous robot scientists like Adam and Eve are promising to create a revolution.



High throughput Experimentation

The outright obvious advantage of using robots to run experiments is that a vastly larger search space can be explored as compared to a human scientist.

Adam studied thousands of strains of yeast carrying out over 6 million measurements during his project cycle. A graduate student who studies fifty strains of yeast in a year would be considered a superstar.

The ability to carry out a large number of experiments both in time and space is a particularly important requirement in discovering new materials.

We know about 120 elements from the periodic table — why can’t we mix each of them in pairs and see what we end up creating? Maybe there is a super alloy here or a new kind of magnet. Perhaps even a superconductor?

Just the binary combination of elements in a 50:50 ratio would require us to create and study 7140 materials. But the catch here is that we would need to study all possible combination ratios! 

No alt text provided for this image

The workflow of Science. The current paradigm typically takes 10 to 20 years to scale a product from conception to commercialization. The emergent AI driven approaches are promising to reduce this time to as little as 5 years. From ‘Inverse molecular design using Machine learning: Generative design for matter engineering’

Efforts are currently underway to solve small subsets of these problems. The High Throughput Experimental Database at the National Renewable Energy Laboratory contains over 60,000 thin films of metal samples. The database contains structural, electrical and optical information of these materials and is accessible to the public.

The Air Force Research Laboratory’s Autonomous Research System (ARES) has been studying the best conditions to rapidly grow carbon nanotubes. By combining mathematical insights as well as genetic algorithms, the system has arrived at its targeted growth rate identifying the conditions most influential in carbon nanotube growth. It got there by running a hundred experiments a day.

With the explosion in deep learning, scientists are combining techniques such as adversarial networks and variational encoders with robotics to create ever more advanced versions of Adams. These systems not only carry out a large number of experiments covering enormous regions of the compositional space, but they also better themselves with each iteration— just as a human scientist would do.

The autonomous scientist has recently made its way into the commercial space with companies such as Kebotix, Atomwise, BenevolentAI and Zymergen devoted to the development of closed loop systems to discover materials, organics and drugs.



However, human scientists are not really in any danger of being replaced by robots. We provide the one thing that advanced automation cannot: Creativity.

 If anything this could mean that the graduate school of the future would be a very different place with every student being able to run hundreds of thousands of experiments at once.

Science would however would change forever inundating us with far too much knowledge for us to know what to do with. 

The world would be unidentifiable with new materials and devices, none of which exist today.

Robots are doing science and that’s a good thing.


Further Reading

  1. Philip Ball, Using artificial intelligence to accelerate materials development, MRS Bulletin, May 2019
  2. Ross King et al, The Automation of Science, Science, 2009
  3. Alan Guzik and Kristin Persson, Materials Acceleration Platform, 201

要查看或添加评论,请登录

Vineeth Venugopal的更多文章

  • The Samsara of Sand

    The Samsara of Sand

    The sand at the Alleppey beach is golden at sunset. ‘Sand is a compelling substance — somehow seductive and sensuous’.

    12 条评论
  • The G-filter: The simple high tech solution to India's water pollution

    The G-filter: The simple high tech solution to India's water pollution

    As Abdul Rajak Khan kneads the clay for his Gravity Filter, the sights and sounds of the village are all around him…

    1 条评论
  • Deep learning with air bubbles

    Deep learning with air bubbles

    Glass identifies faces at the speed of light with zero power 12 July, 2070 7:00 pm I am tired and exhausted as I walk…

    2 条评论

社区洞察

其他会员也浏览了