Toward “On‐Demand” Materials Synthesis and Scientific Discovery through Intelligent Robots

Abstract A Materials Acceleration Operation System (MAOS) is designed, with unique language and compiler architecture. MAOS integrates with virtual reality (VR), collaborative robots, and a reinforcement learning (RL) scheme for autonomous materials synthesis, properties investigations, and self‐optimized quality assurance. After training through VR, MAOS can work independently for labor and intensively reduces the time cost. Under the RL framework, MAOS also inspires the improved nucleation theory, and feedback for the optimal strategy, which can satisfy the demand on both of the CdSe quantum dots (QDs) emission wavelength and size distribution quality. Moreover, it can work well for extensive coverages of inorganic nanomaterials. MAOS frees the experimental researchers out of the tedious labor as well as the extensive exploration of optimal reaction conditions. This work provides a walking example for the “On‐Demand” materials synthesis system, and demonstrates how artificial intelligence technology can reshape traditional materials science research in the future.

MAOS core needs to fetch and store data from the database management system (DBMS) so that the core could work normally. Run python run.py to start MAOS core. The request handler would start on port 9999.
Human-machine interaction MsTimer2 is used to control the time accurately.
• STM32: A STM32 F103 development board is used to control the peristaltic pumps and a stepper motor.

Notes
It is the first version of MAOS; more functionalities are under construction.

Materials property database: This database is built up with online open source such as
Open Quantum Materials Database (OQMD) [1] . Properties such as bandgap, space group id, the magnetic moment of 66993 kinds of inorganic materials are collected.

Regent database:
We sort out 1052 reagent information based on our purchase records.
Information includes purity, concentration, packing size, and price.

Method database:
The method database is constructed with materials synthesis methods detailed with parameters in each step. So far, solution-based methods for various kinds of nanomaterials synthesis including CdSe, CdS, PbS nanocrystal for optoelectronic devices, Au and MoO 2 nanoparticles for plasmonic photothermal applications. An example here is the synthesis method of CsPbBr 3 QDs in Table S1.

S3. The pricing model
Nowadays, the concepts of materials synthesis are not only the reaction of matters, collecting or breaking of chemical bonds, but a deep fusion with the environment, economics, and human life. The evaluation of synthesis experiments should be treated as a hybrid model.
Here, to encourage researchers to approach the design principle of "green chemistry" [2] , a pricing model of MAOS is constructed for evaluating the total cost of each experiment.
It is an overall evaluation of recourse expending including three issues: chemical reagent, materials toxicity, and local prices, forming as The cost of chemical reagent is calculated according to the reagent database, which contains the price information of all purchased reagent in the lab.
is evaluated according to the toxicity and pollution strength of the materials synthesized by the reagents, and the by-products during experiments. This cost is set to encourage researchers to optimize synthesis methods and generate materials with little or no toxicity to human and environment. The identification of is based on Environmental Protection Tax Law [3] promulgated at the 25th meeting of the 12th Standing Committee of the National People's Congress. can be calculated with: Here, is the scale of tax for different kinds of materials with unit mass. For example, Cd-compound is 0.005RMB/kg, and Pb-compound is 0.025RMB/kg, Pxylene is 0.02RMB/kg, details can be found in the Table of Taxable Items and Tax Rates of Environmental Protection Tax [3] . Tax rates are related to the local government of lab.
is evaluated according to the local prices of the lab, which is highly location-dependent.
The cities with high infrastructure and workforce cost will have large . Here we introduce the Global City Index (GCI) [4] to evaluate in following the format: We collected the cost of CsPbBr 3

CHECK_RESOURCES :
checks if the hardware works well and the reagent is enough.
The reaction in charge of communicating with the hardware control interface and controlling the experiment optimizer choose the reaction parameters (action) and learn from the analysis result (reward).

LOAD_REACTION_HISTORY:
load all experiments about this action and update the state of the optimizer.

STORE_RESULT:
15 store the experiment results in the database.

S5. Communication test
Higher data rate and lower delay are eager for collaborative robots and real

Compile
An example of compiling the input experimental formula to machine instruction code is shown in Figure S2. This example shows how MAOS extracts information from input formula, and set up all experimental parameters in configuration file for CdSe QDs synthesis.
The configuration file was generated through designed template. It decides the all adjustable or constant parameters for optimizing the experiment. As shown in Figure. S3, the collaborative robot has four degrees of freedoms. An end grasper is equipped for transfer the reagent containers. Figure. S4 HTC Vive and our AR glass.

VR devices
As shown in Figure. S4, HTC Vive suite, including a VR glass, two gamepads, and two locators are utilized for remote control of the collaborative robot. An AR glass was used to share the live vision in HTC Vive glass.

Chembox
The chamber in Chembox provided a milliliter scale reaction system integrated with multiple functional devices, including temperature, magnetic stir, and in-situ computer vision monitoring. The feeding system combined on reaction chamber of Chembox system (shown in Figure. S5) can fit the requirement to control the injection of both liquid and solid reactants.
The solid sample is stored in a horizontal arranged tube; the push rod can move on the screw rod and motor three control the speed and distance. The push rod is used to push the powder in the tube to the reaction vessel. The whole feeding system can move on the sideway 20 controlled by the motor. A quadrantal half-cylinder-shaped tube fabricated by the 3D printer was used for liquid sample and connected with a peristaltic pump via a rubber tube. All feeding system was controlled by an STM32 chip on top. In this work, we made an upgraded by integrating an automatic weighing system for Chembox system. As shown in Figure. S6, the total mass of added reactant, solution volume, and reaction condition information can be displayed on an LCD interface. Solenoid valves controlled by Arduino chip were utilized to provide vacuum or N2 atmosphere environment for a specific reaction. Temperature calibration. Temperature calibration of chemical flow in heat zone was designed according to Kovalenko and deMello 's work in 2016 [5] . It utilized the decrease of fluorescence intensity of Rhodamine B continuously with increasing the temperature [6] . The

Liquid transfer module
A transfer module was designed as a connection center for different synthesis and characterization modules (Figure. S10). An up-down system made up with a stepping motor 24 and screw rod was assembled for guiding the tube in-out of solution container. The tube guider was manufactured by 3D printer with PLA materials. As shown in Figure. S12, the 3D Printer is used to print machine parts.
Cd Precursor (5mM) was prepared in a 100 mL round-bottom flask. 65 mg of CdO and 3 mL of oleic acid (OA) were added. 47 mL of ODE was then injected through syringe pumps.
Vacuum environment was applied to degas for 30 minutes at room temperature and then Shown in Figure. S13, ten pictures collected original data of all optimized parameter sets which satisfy user demand of . The information of FWHM was indicated by parameter D.

Details of multiple injections of precursors
In the optimization process of the size distribution of nanoparticles, according to the result derived by Clark [8] , the average radius and the standard deviation in ordinary growth In the work present, the scenario to be optimized is to improve the size distribution of the nanoparticles produced with a fixed amount of monomers injected. Inspired by the work of Alivisatos [9] , the optimization process was done by separately injecting the fixed amount in several times (defined as ), each separated by a time interval (defined as ). Thus,

∑ ( )
Thus, right after the -th injection and before the -th injection, is the total reaction time, is the parameter used to scale the reaction speed, which is assumed to be constant. The function is the Heaviside Theta function, which takes only 1 at positive numbers, and zero elsewhere used to resemble the discrete injection. By defining such a , i.e., concentration, since ∫ ⟨ (( ) )⟩ in ordinary growth and ( ) during Ostwald Ripening by Wagner [10] , could be numerically calculated based on the given above with . Thus, if the time was just right after the -th injection and before the -th injection.
By such an expression of , it could be proved that there must be intersections between different schemes. It could be proved from the existence of an intersection in between the indicating possibilities for optimization of the reaction scheme. Note that is definitely not the best option always.

S8. Other Cases of 'On-Demand' Synthesis by MAOS
So far, various kinds of solution-based nanomaterials synthesis for different application situations were covered. CdSe, PbS nanocrystals were synthesized for optoelectronic devices and Au and MoO 2 nanoparticles for photothermal applications.

PbS nanocrystals
We provide PbS nanocrystal for the demand of near-infrared optoelectronic devices. A hotinjection synthesis strategy is used for PbS nanocrystal synthesis [11] . Scheme of automatic synthesis process is shown in Figure.  Finally, change the temperature to 120℃ and keep the N 2 environment by flowing N 2 gas (10 minutes). 0.5ml of TMS was rapidly injected (1ml/s) into Chembox. After 3 minutes reaction, the solution was extracted through cooling, Abs, and PL module.   Au nanoparticle with precisely controlled size and narrow size distribution. The method is based on the reduction of gold chloride with sodium citrate [12] . Pure water is used to adjust the concentration of liquid reagents. CV module here was set to monitor the value and color information of reaction. Absorption of products will be measured with the help of the robot and liquid transfer module.
According to the reference, with the various ratio of Na 3 -citrate : HAuCl 4 , nanoparticles with different size can be synthesized. However, the problem of reproducibility caused by experimental environment sometimes results in a diameter variance, and even worse, poor quality. MAOS needs a random searching process through absorption spectrum and then 33 benchmarked with the data in reference. 15 random parameter set contains different Na 3citrate : HAuCl 4 ratio was first implemented. Figure. S17a shows the spectrum of 9 samples with measurable absorption peak. The samples with the highest absorption peaks (violet and green line) were processed for TEM characterization ( Figure. S17b-c). Thus, the recipe for Au nanoparticles with 5nm and 11nm average size were found, and the mapping of size-ratio was updated based on these experimental results. With this practical mapping, 'On-Demand' synthesis of Au nanoparticles with 5nm, 8nm, 10nm, and 15nm and narrow size distribution were synthesized ( Figure. S17 d-g). Similar to Au nanoparticles, MoO 2 nanoparticle has easily adjustable plasmonic properties and practical synthesis method [13] . By adding different amount of oxidant (H 2 O 2 solution) and reductant (NaBH solution), (Figure. S19) intensity of the plasmonic peak of MoO x (x various from 3 to 2) changes. Figure. S20 shows that most substantial plasmonic peak appears when NaBH:Mo=2. The strong plasmonic signal at near 800nm can satisfy the demand of NIR bio imaging.

Figure. S20
The absorption spectrum of synthesized MoO x nanoparticles.